diff --git a/build.py b/build.py
index 93816a5..dd98c4a 100755
--- a/build.py
+++ b/build.py
@@ -106,6 +106,7 @@ def build_patch(input, output, lang):
         symbols={"LANG": lang},
         include_paths=[Path("src")],
         prelude_file=Path("config.i"),
+        overlap_mode="warn",
     )
 
     if result.exit_code != 0:
@@ -165,7 +166,9 @@ def build_fixed_asset(table, input_file, binary_text_file):
     write_pointers_value_as_binary(pointers, binary_text_file)
 
 
-def build_fixed_to_ptr_asset(table, input_file, binary_text_file, pointers_file, buffer_width=None):
+def build_fixed_to_ptr_asset(
+    table, input_file, binary_text_file, pointers_file, buffer_width=None
+):
     pointers = read_fixed_from_xml(
         input_file, table, formatter=lambda t: t.strip() + "[end]"
     )
@@ -184,7 +187,7 @@ def build_fixed_to_ptr_asset(table, input_file, binary_text_file, pointers_file,
             print(f"{text} is too long ({ptr_len}px , {math.ceil(ptr_len / 8)} tiles)")
 
     text = table.to_text(max_ptr[1].value)
-    print(f"{text} is the largest ({max_length}px ({math.ceil(max_length/8)})")
+    print(f"{text} is the largest ({max_length}px ({math.ceil(max_length / 8)})")
     write_pointers_value_as_binary(pointers, binary_text_file)
 
     write_pointers_addresses_as_binary(
@@ -327,8 +330,7 @@ def add_custom_kernings(text: str, advance: int) -> None:
         known_pairs_to_kern = [
             "Ya",
             "Pa",
-            "Po"
-            "Fa",
+            "PoFa",
             "Fe",
             "Fo",
             "Fu",
@@ -489,6 +491,12 @@ def build_assets(assets):
             menu_table,
         ),
         ("fixed", menu_table, os.path.join(text_root, "items.xml"), "assets/items.dat"),
+        (
+            "fixed",
+            menu_table,
+            os.path.join(text_root, "items_unleashed.xml"),
+            "assets/items_unleashed.dat",
+        ),
         ("fixed", menu_table, os.path.join(text_root, "magic.xml"), "assets/magic.dat"),
         (
             "fixed",
@@ -528,7 +536,7 @@ def build_assets(assets):
             os.path.join(text_root, "monsters_long.xml"),
             "assets/monsters_long.dat",
             "assets/monsters_long.ptr",
-            80
+            80,
         ),
         (
             "nullterminated",
diff --git a/config.i b/config.i
index a3e296a..a185474 100644
--- a/config.i
+++ b/config.i
@@ -13,6 +13,7 @@ ENABLE_BUTTON_DISPLAY := 1
 BATTLE_CMD_VWF := 1
 BATTLE_NAMES_VWF := 1
 BATTLE_MONSTERS_VWF := 1
+BATTLE_ITEMS_VWF := 1
 INVENTORY_ROLLING_BUFFER := 1
 TREASURE_INVENTORY_ROLLING := 1  ; WIP, treasure rolling-buffer scaffolding only
 TREASURE_DEBUG_ALWAYS_DROP := 1  ; force every win to drop an item (testing)
diff --git a/ff4.s b/ff4.s
index 45b4942..bf319ae 100644
--- a/ff4.s
+++ b/ff4.s
@@ -53,42 +53,48 @@ conditional_bg1_vofs := 0x208000
 dialog_bank_ptr_base = 0x218000
 
 
-*=0xFFC0
+.alloc at 0x00FFC0 {
     ; patch snes cartridge type
-    ; original PCB: SHVC-1A3B
-    ; target PCB: SHVC-1A5B
+    ; original PCB: SHVC-1A3B  ;  target PCB: SHVC-1A5B
     .ascii "Final Fantasy IV     "
+}
 
-;FFD5 20H / 30H Map Mode
-
-*=0xFFD6
+.alloc at 0x00FFD6 {
+    ; FFD5 20H / 30H Map Mode
     .db 0x02  ; Cartridge Type
     .db 0x0B  ; ~ 0BH ROM Size
     .db 0x07  ; RAM Size
+}
+
 .if ENABLE_BRK_HANDLER {
-    *=0x00FFE0
-; JML trampoline in vector-table padding  ; native/emu BRK vectors point here.
-    jmp.l brk_handler
-    *=0x00FFE6
-    .dw 0xFFE0
-    *=0x00FFFE
-    .dw 0xFFE0
+    ; JML trampoline in vector-table padding; native/emu BRK vectors point here.
+    .alloc at 0x00FFE0 {
+        jmp.l brk_handler
+    }
+    .alloc at 0x00FFE6 {
+        .dw 0xFFE0
+    }
+    .alloc at 0x00FFFE {
+        .dw 0xFFE0
+    }
 }
 
 
-*=0x008031
-    ; déroutage pour ajouter le splash screen
-.if ENABLE_INTRO {
-    jsr.l start_splash_screen
-} else {
-    jsr.l clear_ram
+; déroutage pour ajouter le splash screen
+.alloc at 0x008031 {
+    .if ENABLE_INTRO {
+        jsr.l start_splash_screen
+    } else {
+        jsr.l clear_ram
+    }
 }
 
 
-*=0x00B463
-    ; déroutage pour utiliser la vwf dans les dialogues.
+; déroutage pour utiliser la vwf dans les dialogues.
+.alloc at 0x00B463 {
     jsr.l vwfstart
     rts
+}
 
 ; ============================================================================
 ; Bank-20 relocated region.
@@ -129,8 +135,14 @@ _cond_skip_bg1vofs:
 
 clear_ram:
 """
-Clear the dialog VWF tile buffer at $702000-$706FFF (16-bit zeroes) after letting the boot ROM init at
-$15C9AA.
+Clear the dialog VWF tile buffer + engine scratch at $702000-$7070FF
+(includes VWF_CONFIG_BASE, VWF_CHR_DIRTY / DIRTY_B, VWF_CALLER_CTX, and
+the secondary descriptor fields) after letting the boot ROM init at
+$15C9AA. Range was $5000 bytes pre-secondary-descriptor  ; bumped to
+$5100 so the new dirty / vram_word / byte_count / src_offset bytes
+land zero on cold boot instead of inheriting random SRAM and
+triggering a bogus secondary flush on the very first NMI (which trashed
+the save-selection sprite CHR).
 """
 
 
@@ -142,7 +154,7 @@ $15C9AA.
 _loop:
     sta.l 0x702000, x
     inx
-    cpx.w #0x5000
+    cpx.w #0x5100
     bne _loop
     }
     rtl
@@ -167,6 +179,30 @@ Output: X = offset into ItemName table.
     rtl
 
 
+multiply_item_index_17:
+"""
+Relocated multiply-by-ITEM_UNLEASHED_RECORD_SIZE for the items_unleashed
+name offset. Called from $019023 via JSL when the field menu is
+wired to the 17-byte assets_items_unleashed_dat table.
+Input: $43 = item ID (16-bit mode active).
+Output: X = offset into ItemName table.
+"""
+
+
+; ITEM_UNLEASHED_RECORD_SIZE = 17 = (id << 4) + id.
+    lda 0x43
+    pha
+    asl
+    asl
+    asl
+    asl  ; * 16
+    clc
+    adc 0x01, s  ; * 16 + id = * 17
+    tax
+    pla  ; balance stack
+    rtl
+
+
 multiply_by_12:
 """A: value to multiply  ; returns A*12 in A."""
     php
@@ -184,6 +220,29 @@ multiply_by_12:
     rtl
 
 
+multiply_by_17:
+"""
+A: value to multiply  ; returns A*17 in A. Mirror of multiply_by_12
+sized for the 17-byte assets_items_unleashed_dat stride.
+"""
+
+
+    php
+    rep #0x20
+    and.w #0x00FF
+    pha
+    asl
+    asl
+    asl
+    asl
+    clc
+    adc 0x01, s  ; * 16 + value = * 17
+    sta 0x01, s
+    pla
+    plp
+    rtl
+
+
 brk_handler:
 """
 BRK trap: mask interrupts, disable NMI, fetch the BRK signature byte
@@ -223,7 +282,6 @@ signature byte sits at PB:(PC - 1).
 ; and matches the legacy `*=0x208000` chain so .import modules without
 ; their own `*=` directive land in bank-20 as expected.
 
-*=0x208100
     ; --- Imported modules ---------------------------------------------------
 
 .import "libmz"
@@ -253,6 +311,7 @@ signature byte sits at PB:(PC - 1).
 }
 
 .import "ingame/places_names_window"
+.import "ingame/items_menu_vwf"
 .import "menus/system_menus_text"
 .import "dakuten"
 .import "menus/start_screen_text"
@@ -265,6 +324,7 @@ signature byte sits at PB:(PC - 1).
 .if INVENTORY_ROLLING_BUFFER {
     .import "ingame/init_bg_scroll_hdma"
     .include "src/ingame/inventory_rolling.s"
+    .include "src/lib/rolling_inventory_engine.s"
 }
 
 .if TREASURE_INVENTORY_ROLLING {
@@ -275,6 +335,7 @@ signature byte sits at PB:(PC - 1).
 
 ; --- Binary text assets -------------------------------------------------
 
+
 .incbin "assets/attack_names.ptr"
 .incbin "assets/attack_names.dat"
 .incbin "assets/monsters_long.ptr"
@@ -310,4 +371,12 @@ signature byte sits at PB:(PC - 1).
     stx 0xb3
     jmp.w 0xED96
 }
-;end
+
+; Park the 17-byte-stride items_unleashed.dat in an empty bank so the
+; full 4352-byte table fits without crossing a LoROM bank boundary
+; (which would otherwise leave the upper half of the table at
+; $21:0xxx, an address LoROM does not map back to ROM data).
+
+.alloc at 0x238000 {
+    .incbin "assets/items_unleashed.dat"
+}
diff --git a/plans/battle_items_vwf.md b/plans/battle_items_vwf.md
new file mode 100644
index 0000000..ebe9bb3
--- /dev/null
+++ b/plans/battle_items_vwf.md
@@ -0,0 +1,117 @@
+# Battle items VWF
+
+## What
+
+Replace the fixed-width 12-character item-name rendering in the battle
+inventory with the existing 8x8 VWF blitter (`battle_render.display_char`).
+Lets proportionally-spaced French names fit the 15-tile slot budget that
+the rolling inventory already reserves, removes hard truncation, frees
+the per-glyph padding wasted on narrow characters.
+
+## Why
+
+Vanilla `DrawInventoryItemText` packs item names as 12 fixed tile IDs
+into a 48-byte buffer. Our rolling inventory expanded to 60 bytes (15
+tiles per row × 2 bytes), but the renderer still emits one tile-id per
+character, so a 12-character French name like `Médaille d'Or` either
+overflows or truncates. VWF rendering pulls the same name through
+`battle_render.display_char`, allocating CHR dynamically and packing
+proportional glyphs into 15 tiles of horizontal pixels (~120 px).
+
+## Boundaries
+
+- In scope: battle inventory (the in-battle item menu).
+- In scope: the rolling-buffer slot refresh hook (re-render on
+  scroll edge).
+- Out of scope: equipped-items buffer ($9A00). Equip menu rendering
+  stays fixed-width for this PR; can follow up.
+- Out of scope: item descriptions (already VWF via small_vwf).
+- Out of scope: field inventory VWF. Same approach, different surface.
+
+## Approach
+
+1. **Reserve a battle_render region for inventory**. Current regions
+   (`src/battle/message.s:143-150`):
+   ```
+   0x00-0x3F messages
+   0x40-0x7F monster names
+   0x80-0xAF char names
+   0xB0-0xEF commands
+   ```
+   Add `0xF0-0xFF` (16 tiles = 512B CHR) for inventory item name. With
+   6 ring slots and 15 tiles per name, the ring needs ~90 tiles. Either
+   extend the CHR window past $BDE0 (free runway up to roughly
+   $BFE0-ish per the audit) or reuse the commands region during the
+   item-menu mode (commands hide when inventory is open).
+2. **Patch `DrawInventoryItemText`** at `$02:9FA4`. Replace the
+   per-character tile-id copy with a `battle_render.init_inventory`
+   + per-byte `display_char` loop, then write the resulting tile IDs
+   into the existing 60-byte slot at `$97A6 + slot * 60`.
+3. **Quantity digits**. Stay fixed-width for now (the rightmost 2
+   tiles per row in the slot layout). Avoids a special-case digit
+   path in display_char.
+4. **Trash icon**. Existing 2x2 fixed glyph stays. Item id $FF skips
+   the VWF render path.
+5. **Tilemap layout unchanged**. The slot still consumes 30 tiles
+   (15 × 2 rows), still gets DMA'd to VRAM tilemap by the existing
+   rolling-buffer transfer. Only the CHR pointed at by the tile IDs
+   changes.
+
+## Files
+
+- `src/battle/items_patches.s` — hook `DrawInventoryItemText` text
+  copy loop.
+- `src/battle/message.s` — add inventory region constants +
+  `init_inventory` entry.
+- `src/battle/inventory_rolling.s` — slot-refresh callsite. Already
+  pre-renders on swap / scroll; VWF render lands at the same site.
+- `tests/test_battle_init.py` — extend with an "open battle
+  inventory" frame that captures the VWF rendered names.
+
+## Tests
+
+- Battle init smoke unchanged (no item menu open).
+- New `tests/test_battle_inventory_vwf.py`: load a battle-with-items
+  kss, open the item menu, capture screen golden showing
+  VWF-rendered names.
+- `tests/_profile/profile_battle_vwf.py` exercised with the inventory
+  open; confirm `_display_char` calls fit within budget on edge
+  frames (re-render is one slot, ~15 chars × 4.4k cy = 66k cy ≈ 18%
+  of frame).
+
+## Risks
+
+- **CHR overflow**: 90 tiles for the ring needs $A000-base allocation
+  past current region. Audit the $BDE0-? runway carefully before
+  expanding.
+- **Tile-id collision with commands**: if commands region 0xB0-0xEF
+  is the chosen reuse target, items render must wait for cmd window
+  to be hidden. The existing battle inventory open path already
+  hides the cmd window via the slice-2 gates, but verify.
+- **Edge-slot render time**: VWF blit of one item name on scroll
+  edge must fit in ~1 vblank. Item names are ~12-15 chars × 4.4k cy
+  = 53-66k cy ≈ 15-18% of frame budget. Headroom OK.
+- **dakuten/special chars**: item names with accents flow through
+  `battle_display_dakuten_char`. Already handled by the existing VWF
+  path. Verify on `Médaille` etc.
+
+## Phasing
+
+| phase | scope | size |
+|---|---|---|
+| 1 | Region reservation + region-init helper | ~50 lines asm |
+| 2 | `DrawInventoryItemText` VWF hook + slot write | ~150 lines asm |
+| 3 | Rolling-buffer slot-refresh wiring | ~80 lines asm |
+| 4 | Goldens + smoke test for battle inventory | tests |
+| 5 | Polish: quantity-digit alignment, trash icon | small |
+
+Each phase commits independently. PR can land as either one chunk or
+phase-by-phase depending on review cadence.
+
+## Crack
+
+The 90-tile CHR budget is the load-bearing assumption. If the
+$A000-$BFFF window proves too tight, fall back to reusing commands
+region 0xB0-0xEF during item-menu mode and round-trip the cmd window
+on close. Adds a tilemap clear on transition but keeps the CHR
+footprint where it is.
diff --git a/plans/menu_rewrite.md b/plans/menu_rewrite.md
new file mode 100644
index 0000000..baa4b31
--- /dev/null
+++ b/plans/menu_rewrite.md
@@ -0,0 +1,140 @@
+# Menu system rewrite
+
+## Why
+
+FF4's menu/UI layer accumulated technical debt that costs disproportionate
+effort on every new feature:
+
+- 4 BG3 CHR bases that flip per context (menu `$4000`, town `$4000`-small,
+  dialog `$C000`, battle `$5000`).
+- 4 staging buffers (`$0774` text, `$7E:D600` BG3, `$7E:B600` BG1,
+  `$7E:C600` BG4).
+- Tilemap bases shift per BG mode.
+- Vanilla `$EB=1` NMI dispatch interleaved with custom rolling-engine hooks.
+- Tile-id semantics flip 8-bit / 9-bit by mode.
+- Per-menu HDMA channels (5, 6, 4) each with their own header/footer/signal
+  hook trio just to anchor the window border while items scroll.
+- Save/restore range depends on caller (`$1300` widened to `$2000` then
+  bumped again for the VWF leak).
+
+Every new menu feature is an archaeology dig. The current rolling engine
+landed in 18 commits and still doesn't cover the key-item picker because
+that runs as an overlay on the active map and the existing engine's
+assumptions don't hold there.
+
+## Target architecture (FF6-derived)
+
+Both FF5 and FF6 simplify in two key ways FF4 didn't pursue:
+
+1. **Window borders and item text live on different BG layers.** FF6 puts
+   the item list on BG1 (HDMA-scrollable), window frames on BG2 (static),
+   cursor sprites on BG3. The "anchor the bottom border while items
+   scroll" gymnastics dissolve — borders don't move.
+2. **HDMA scroll uses a mostly-static ROM table** with one runtime patch
+   for the scroll bias. FF6's `LoadItemBG1VScrollHDMATbl`
+   (`ff6decomp/src/menu/item.asm:270`) copies a ROM-resident table into
+   WRAM and `ADC`s the current scroll position into the variable slice.
+   No per-frame buffer_pos / buffer_slots modular math.
+
+### Proposed VRAM map
+
+```
+$0000-$2FFF  BG1 + BG2 CHR (mode-1 4bpp, shared map tileset)
+$3000-$3FFF  BG1 tilemap (item text — HDMA-scrollable layer)
+$4000-$4FFF  ??? (consider relocating sprite CHR here once layout settles)
+$5000-$5FFF  BG2 tilemap (window borders — static)
+$6000-$7FFF  reserved
+$8000-$BFFF  sprite CHR
+$C000-$DFFF  BG3 CHR (dialog VWF + description font — shared by all
+             BG3-text contexts incl. picker)
+$E000-$EFFF  BG3 tilemap (description / overlay text)
+```
+
+Single CHR window for BG3-text across menu, dialog, picker. No flipping
+BG34NBA per context.
+
+### HDMA strategy (FF6-mirrored)
+
+Per menu:
+- Static ROM table `<menu>_HdmaTable` with pre-computed `count + scroll`
+  entries for every row of the list. Hand-tuned per menu's vertical layout.
+- One runtime subroutine `LoadHdmaTable<menu>` copies ROM table → WRAM
+  shadow (e.g. `$7E:9800`), `ADC`s the variable slice with current
+  scroll_pos. ~10 instructions.
+- HDMA channel 5 hardcoded to scroll BG1VOFS. Only one channel needed;
+  no `header_hook` / `footer_hook` / `signal_hook` dispatch.
+
+### Window draw
+
+Single `DrawWindow` writes border tiles to whichever BG staging buffer
+is selected. Caller selects BG2 for window frames, BG1 for body content.
+No animation, no HDMA window-position channels. Window appears
+instantly on tilemap blit.
+
+## Function replacement list
+
+Replace wholesale (vanilla site becomes `JSR thunk; RTS` in bank-01,
+thunk JSL's bank-20 reimpl). Old vanilla code becomes unreachable.
+
+| Vanilla | Address | Replaces | Blocks |
+|---|---|---|---|
+| `DrawWindow` | `$01:80D9` | Custom window-frame blitter writing to selected BG | BG layer split |
+| `UpdateItemText` | `$00:B229` | Custom VWF render of N visible rows into BG1 tilemap | Picker integration |
+| `DrawItemSlot` inner | `$01:A1ED` | Custom row writer matching new tilemap layout | Per-menu render hooks |
+| `ShowItemWindow` | `$00:AF4D` | Custom picker driver routed through rolling engine | Standalone picker |
+| `TfrBG1/2/3/4TilesVblank` | `$01:9420-9447` | Unified NMI DMA driver (one entry, dirty-mask dispatch) | 4-trampoline spaghetti |
+| Main field-menu loop | `$01:9F7B+` | Custom dispatcher running rolling engine + new draw stack | Vanilla input FSM |
+
+Routines that stay (layer-agnostic, no need to touch):
+- `InitItemList` (`$01:B2D3`) — list filter
+- `DrawItemCursors` (`$01:A105`) — sprite hand
+- `CheckCanUseItem` (`$01:A25D`) — palette decision
+
+## Migration order
+
+Each step is one commit, parking-lot tested via screenshot goldens.
+
+1. **`DrawWindow` reimpl in bank-20** (mirrors vanilla output exactly).
+   No call-site changes. Verify byte-for-byte identical render via test.
+2. **Switch own trampolines** (`draw_window_trampoline`,
+   `_treasure_draw_inventory_window`, `_menu_draw_inventory_window`,
+   `_drops_draw_window`) to the new entry. Vanilla `$01:80D9` still alive
+   for vanilla callers (event dialogs, etc).
+3. **Hijack `$01:80D9`** with `JSR _drawwindow_thunk; RTS`. Vanilla code
+   becomes dead.
+4. **Per menu: redirect window draw to BG2 staging.** Wrap each menu's
+   draw-window call with `SelectBG2 → DrawWindow → SelectBG1`. Border
+   moves from BG1 to BG2.
+5. **Per menu: delete HDMA header/footer hooks.** Now that border is on
+   BG2 (static), BG1 HDMA only needs per-row scroll entries. Collapse
+   `update_X_scroll_hdma` to FF6-style static-table + scroll-bias patch.
+6. **`UpdateItemText` reimpl.** Replaces vanilla item-list rendering.
+   Routes through small-VWF directly.
+7. **`DrawItemSlot` reimpl.** Same pattern; replaces per-slot vanilla
+   render.
+8. **Picker wire-up.** Now trivial — picker is just `DrawWindow` (border
+   on BG2) + `UpdateItemText` (VWF on BG1 region) + rolling engine for
+   scroll. No more "town context vs menu context" branches.
+9. **`TfrBG*TilesVblank` unification.** One vblank DMA driver dispatched
+   by a dirty mask. Replaces the 4 vanilla trampolines + each menu's
+   per-BG transfer-pending shadow.
+10. **Main field-menu loop reimpl.** Last large vanilla function to
+    delete. Drives rolling engine + new draw stack from a single
+    dispatcher.
+
+## Acceptance gates
+
+- Every commit keeps the 143/19/1 baseline (no new fails).
+- Screenshot goldens flag any pixel diff for review — re-record only
+  when the diff is intentional.
+- After step 5, expect to record new field-menu screenshot goldens
+  (border-on-BG2 visual is identical but tilemap snapshot will differ).
+- After step 10, delete `src/lib/rolling_buffer.s` macro stubs +
+  bank01_trampolines entries that point at unreachable vanilla addresses.
+
+## Out of scope
+
+- World map / airship / battle inventory (different BG modes; pattern
+  port-able later once the field-menu pipeline is proven).
+- Mode-7 transitions.
+- HDMA-driven palette/scroll effects outside the inventory list.
diff --git a/plans/rolling_inventory_engine.md b/plans/rolling_inventory_engine.md
new file mode 100644
index 0000000..a767591
--- /dev/null
+++ b/plans/rolling_inventory_engine.md
@@ -0,0 +1,276 @@
+# Rolling Inventory Engine — Unified Bank-20 Design
+
+Status: draft. Phase 1 = dynamic VWF render preserved. ROM pre-rendered slots
+deferred to a later phase.
+
+## Goal
+
+One engine, many menus. Today `inventory_rolling.s`, `treasure_rolling.s`,
+`drops_rolling.s`, `key_item_picker.s`, and `shop.s` each duplicate slot
+math, scroll animation, HDMA setup, and VWF render orchestration. Drops +
+treasure currently share a single VWF CHR region and clobber each other.
+
+Phase 1 collapses them onto a shared engine living in bank-20, driven by a
+per-instance state struct. Each menu owns its own struct instance at a fixed
+WRAM base. The engine never holds inventory-specific knowledge — render and
+draw-window hooks are supplied per instance via the struct.
+
+## Non-goals (Phase 1)
+
+- Pre-rendering slot tiles into ROM and DMAing them in. Comes later.
+- Touching the battle inventories. They reuse leaf helpers at most.
+- Changing the visible layout, kerning, or tile budget per menu.
+
+## Architecture
+
+```
+bank-01 caller (e.g. items_menu, treasure, drops, key_items, shop)
+        |
+        | JSL with state struct pointer in (e.g.) X
+        v
+bank-20 rolling_inventory engine
+        |
+        | reads state.* fields
+        | calls state.render_slot_hook / state.draw_window_hook
+        v
+per-instance hooks back in bank-01 (small, menu-specific)
+```
+
+Engine = scroll state machine + slot index math + VWF region setup +
+HDMA bookkeeping. Hooks = "how do I draw the window chrome" and "how do I
+render slot N into the active VWF region for this menu". Hooks stay in
+caller's bank.
+
+## Extended state struct
+
+`RollingBufferState` today is 12 bytes of engine scratch. Phase 1 extends it
+with instance config + VWF region fields so the engine reads everything off
+the struct pointer. No globals like `VWF_CONFIG_BASE` floating around.
+
+```
+.struct RollingBufferState {
+    ; --- engine scratch (12 bytes, unchanged) ---
+    byte top_row
+    byte buffer_pos
+    byte edge_row
+    byte slot_index
+    word base_scroll
+    byte hdma_enable
+    byte _pad
+    byte scroll_state
+    byte scroll_remaining
+    byte scroll_direction
+    byte transfer_pending
+    word scroll_anim_offset
+    byte hdma_copy_pending
+
+    ; --- instance config ---
+    byte visible_rows           ; e.g. 5 field items, 3 shop
+    byte slot_height_tiles      ; rows of CHR per slot (usually 2)
+    long item_list_ptr          ; far ptr to inventory array
+    byte item_count             ; live count (decremented on use/drop)
+    byte hdma_channel           ; statically assigned per instance
+
+    ; --- VWF region (pointer to standalone VwfRegionConfig, see below) ---
+    long vwf_cfg_ptr
+
+    ; --- hooks (far ptrs, bank-01 routines) ---
+    long fn_draw_window         ; draw frame, headers, static chrome (init only)
+    long fn_ensure_hdma         ; install instance HDMA channel
+    long fn_render_slot         ; render slot at A=index into vwf_stage_ptr
+
+    ; --- dirty tracking ---
+    byte dirty_mask             ; bit per visible row, max 8 rows
+}
+```
+
+Size: ~31 bytes per instance. Five instances = 155 bytes WRAM. Plus 11
+bytes per `VwfConfig` (trimmed existing struct from `src/vwf_state.i:74`,
+one per rolling instance + any extras for battle).
+
+### Trim + reuse existing `VwfConfig` struct
+
+`src/vwf_state.i:74` already defines `.struct VwfConfig` (14 bytes) as a
+forward-compat reservation. Nobody reads it yet ; `VWF_CONFIG_BASE` is one
+SRAM slot at `$70:7080`. That single slot IS the clobber bug: coexisting
+instances (drops + treasure) need disjoint configs.
+
+`VwfConfig` also carries dead weight: `font_ptr` / `font_bank` /
+`kerning_ptr` / `kerning_bank`. We have one font globally and kerning is
+already a `flags` bit. Drop those 6 bytes.
+
+Trimmed struct (11 bytes):
+
+```
+.struct VwfConfig {
+    word tile_id_base    ; first tile_id this renderer owns
+    byte slot_budget     ; tile-budget clamp ; $FF = none
+    word tilemap_base    ; WRAM dest base (bank-$7E implied)
+    byte palette_byte    ; default palette / attr
+    byte flags           ; bit 0 = kerning, bit 1 = priority OR ; rest reserved
+    word chr_vram_word   ; VRAM word addr for DMA dest
+    word chr_byte_count  ; DMA byte count
+}
+```
+
+Plan: stop inventing a parallel struct. Trim + consume `VwfConfig`:
+
+1. Trim dead font/kerning ptr fields out of `src/vwf_state.i`.
+2. Allocate per-instance `VwfConfig` slots at fixed SRAM bases (`$70:7080`,
+   `$70:708C`, `$70:7098`, ...). One per rolling instance + any battle
+   consumer.
+3. Migrate `src/vwf.s` / `src/small_vwf/render.s` to read `VwfConfig` fields
+   instead of hardcoded tile_id base + tilemap dest. Font ptr stays the
+   single global it always was.
+
+`RollingBufferState.vwf_cfg_ptr` (`long`) points at the instance's
+`VwfConfig`. Battle reuses the same struct.
+
+Resolved design points:
+
+- **No `fn_format_qty`.** Render hook owns layout. Key items hook writes name
+  only; field items hook writes name + qty. Engine never branches on it.
+- **No bank fields.** `item_list_ptr` and `vwf_stage_ptr` are `long`. Removes
+  "did you set the bank" footgun.
+- **HDMA channel = static struct field, not free-list.** Coexistence map is
+  known at design time (only drops + treasure overlap), so allocation is a
+  static table: drops gets ch 3, everyone else ch 2. Verify against current
+  `src/ingame/init_bg_scroll_hdma.s` before locking numbers.
+- **Shop qty overlay = caller's problem.** Shop calls
+  `rolling_engine_invalidate_slot` on qty change; its render hook draws the
+  overlay. Engine stays slot-agnostic.
+- **`fn_draw_window` lives in struct** even though only called at init. Keeps
+  init signature clean (just `X = state ptr, Y = top_row`); 3 bytes / instance
+  is noise.
+- **`dirty_mask` capped at 8 rows.** No current menu needs more. Locked.
+
+## WRAM instance bases (proposal)
+
+| Instance      | Base       | Notes                                    |
+|---------------|------------|------------------------------------------|
+| field items   | `$7E:1BA8` | already there                            |
+| treasure inv  | `$7E:1BD0` | already there                            |
+| drops         | `$7E:1C00` | new; owns disjoint VWF region            |
+| key items     | `$7E:1C30` | already at 1BF0, may relocate            |
+| shop          | `$7E:1C60` | new                                      |
+
+Field menu + shop never coexist with treasure/drops, but drops + treasure DO
+coexist on the same screen, so their `vwf_chr_base` MUST differ. That is the
+bug we fix.
+
+## Engine entry points (bank-20)
+
+All entry points take `X = instance struct ptr` (16-bit, in WRAM, bank fixed
+to `$7E`). JSL from bank-01 callers.
+
+```
+; Cold init. Zero engine scratch, latch instance config from caller-prefilled
+; fields, draw window, render initial visible slots.
+;   in : X = state ptr
+;        Y = initial top_row (usually 0)
+;   out: VWF region populated, HDMA armed
+rolling_engine_init:
+
+; Per-frame tick. Read pad, advance scroll state machine, schedule VWF
+; renders for newly-revealed rows, update HDMA scratch.
+;   in : X = state ptr
+;        A.b = pad-down bits this frame
+;   out: state.dirty_mask reflects rows needing redraw
+rolling_engine_tick:
+
+; NMI-time flush. Walk dirty_mask, DMA VWF stage -> VRAM, write tilemap.
+; MUST be called inside force-blank or HDMA gap window.
+;   in : X = state ptr
+rolling_engine_vblank_flush:
+
+; Cursor / selection helpers — thin wrappers over slot_index math.
+rolling_engine_cursor_up:
+rolling_engine_cursor_down:
+
+; Mutation API for callers that add / remove items mid-menu.
+;   in : X = state ptr
+;        A = slot index being mutated
+; Marks affected rows dirty; engine re-renders next tick.
+rolling_engine_invalidate_slot:
+rolling_engine_invalidate_all:
+
+; Teardown. Tear down HDMA channel, leave state struct intact for re-entry.
+rolling_engine_shutdown:
+```
+
+Internally the engine reuses the existing `engine_*` macros from
+`src/lib/rolling_buffer.s` (scroll prepare / start / update / finish). Those
+macros become the engine's private kernel; bank-01 callers stop including
+them directly.
+
+## Caller skeleton (bank-01)
+
+Each menu shrinks to: declare instance, fill config fields, JSL init,
+JSL tick from main loop, JSL flush from NMI, supply 3-4 hook routines.
+
+```
+items_menu_open:
+    ldx.w #field_items_state
+    lda.w #visible_rows_5 : sta.w 0,X RollingBufferState.visible_rows
+    ; ... fill remaining config + hooks once ...
+    ldy.w #0
+    jsl rolling_engine_init
+    rts
+
+items_menu_render_slot:        ; A = slot index, X = state ptr
+    ; existing VWF render code, writes to state.vwf_stage_ptr
+    rtl
+```
+
+Hooks are `.long` because engine lives in bank-20 and hooks in bank-01.
+Engine does `jsl [state + fn_render_slot]` style indirection (via a small
+trampoline since 65816 has no JSL-indirect).
+
+## Migration plan — single branch, multi-commit, squash at merge
+
+Single feature branch. Each commit green at HEAD. TDD per commit (failing
+test first, commit when green). Squash to one commit at merge time per
+`feedback_no_rebase_published_branches`.
+
+0. **Consume existing `VwfConfig`.** Wire `src/vwf.s` and
+   `src/small_vwf/render.s` to read from a `VwfConfig*` instead of
+   hardcoded font ptr + tile_id base. Allocate one `VwfConfig` slot per
+   coexisting instance (drops + treasure get distinct slots). Legacy
+   callers point at the original `$70:7080` slot ; un-migrated paths
+   unchanged. Clobber bug fixed at VWF layer.
+1. **Engine skeleton in bank-20** + extended struct + smoke harness driving
+   a fake inventory. No callers migrated. Existing `engine_*` macros become
+   engine's private kernel. Engine reads `state.vwf_cfg_ptr` and passes it
+   straight through to VWF entry points.
+2. **Port field items.** Drop duplicated scroll state machine from
+   `inventory_rolling.s`. Goldens: `tests/goldens/field_items/*` unchanged.
+3. **Port shop.** Second consumer flushes single-instance assumptions.
+4. **Port key items**, enable actual rolling (today it doesn't). New
+   goldens for rolling behaviour.
+5. **Split drops + treasure VWF regions.** Each gets own `vwf_chr_base` and
+   `vwf_dst_tile_id`. New combined golden verifies no clobber.
+6. **Delete duplication.** Remove dead code from `inventory_rolling.s` /
+   `drops_rolling.s` / `treasure_rolling.s` once all callers are on engine.
+
+## Battle (out of scope, future note)
+
+Battle inventory menus have a different tile budget, OAM-driven cursor, and
+tighter NMI window. They can reuse:
+
+- the slot math helpers (which row is visible, where does it map in CHR),
+- the VWF stage -> VRAM DMA helper,
+- the dirty-mask flush.
+
+They cannot reuse the field scroll-animation state machine as-is. Punt to a
+later design once the field side stabilises.
+
+## Verification gates before phase 6
+
+- xdds disassembly of patched ROM at each engine entry point: confirm
+  bank-20 location, MX flags, JSL targets resolve.
+- Goldens green for field items, shop, key items, drops+treasure combined.
+- Mesen-S smoke run: open each menu, scroll past wrap on both directions,
+  add/remove items mid-menu, verify dirty_mask handling.
+- WRAM check: instance bases do not overlap and stay inside the existing
+  `$7E:1Bxx` rolling region (extend allocation if needed; document in
+  `sram_management.s`).
diff --git a/src/bank20.i b/src/bank20.i
new file mode 100644
index 0000000..0a917ea
--- /dev/null
+++ b/src/bank20.i
@@ -0,0 +1,11 @@
+"""Shared bank-20 relocated-code pool declaration.
+
+Every module that wraps its body in `.alloc <name> in bank20_reloc { ... }`
+must include this header so the pool is visible at the module's
+codegen pass. Pool decls are idempotent across modules — the
+linker dedupes identical ranges via `_merge_one_pool_decl`."""
+
+.pool bank20_reloc {
+    range 0x208000 0x20FFFF
+    strategy order
+}
diff --git a/src/battle/commands_reloc.s b/src/battle/commands_reloc.s
index c4cda17..2bb0861 100644
--- a/src/battle/commands_reloc.s
+++ b/src/battle/commands_reloc.s
@@ -1,4 +1,6 @@
 """
+.include "../bank20.i"
+
 Relocated battle-command-list renderer (`draw_command_list_for_character`) plus its private command-build
 loop, called by the patched bank-$02 hook.
 """
@@ -13,172 +15,175 @@ loop, called by the patched bank-$02 hook.
 
 mult8_far := 0x2855c
 
-.if BATTLE_CMD_VWF {
-    command_length = 6
-} else {
-    command_length = 10
-}
 
-draw_command_list_for_character:
-"""
-Public RTL entry: render the per-character battle command list for the active character.
-Dirty-bit gating moved up to the bank-02 thunk
-(`_draw_battle_command_window_relocated`) which clears CMD_DIRTY_BIT
-before calling here  ; gating twice would leave the inner buffer
-filled with $00FF blanks. Inventory-open short-circuit stays.
-"""
-; Skip command rendering if inventory is active (bit 2 of $4A)
-; This prevents format buffer ($74FD) conflicts with inventory code
-    lda.l 0x7E004A
-    and #0x04
-    bne _skip_commands
-
-    lda 0x1822  ; selected character slot
-    sta 0x1816
-    phx
-    jsr.w _draw_command_list_for_character
-    plx
-
-_skip_commands:
-    rtl
-
-_draw_command_list_for_character:
-{
-    stz.w 0x1817
-    ldy #0x74FD  ; keep the pointer in Y this will allow to run _draw_single_command multiple times
-; to build {tile_flag}cmd1\n{tile_flag}cmd2\n{tile_flag}cmd3\n{tile_flag}cmd4\n{tile_flag}cmd5\0
-; to issue a single draw text call to the command list region.
-    lda 0x1817  ; battle command slot
-    sta 0x26
-    lda.b #command_length * 2
-    sta 0x28
-    jsr.l mult8_far  ; command_id * (command_length * 2)
-
-
-    lda 0x1817  ; slot ID
-    asl
-    asl
-    tax
-    stx 0x00
-
-;lda     0x1816       ; character id
-;asl
-;tax
-
-;rep #0x20
-;ldx.w #command_buffer_ptr - 10
-;lda.l     0x16fead,x ;cmd_text_buf_ptrs,x
-;stx     0xef52 ; destination ptr ?
-
-_loop:
-    tdc
-    lda 0x1817  ; slot ID
-    asl
-    asl
-    tax
-    stx 0x00
-
-    lda 0x1816  ; character id
-
-    rep #0x20
-    asl
-    tax
-    lda.l 0x16feb7, x
-    clc
-    adc.b 0x00
-    tax
-    sep #0x20
-
-    jsr.w _draw_single_command
-    inc 0x1817
-    lda 0x1817
-    cmp #5
-    beq _exit
-    lda #1
-    sta.w 0x0000, y
-    iny
-    bra _loop
-_exit:
-
-    tdc
-    sta.w 0x0000, y
-
-    sta 0xef55
-
-    ldx.w #command_buffer_ptr
-    stx 0xef52  ; destination
-
-    rep #0x20
-
-; Fill the window with 0x00ff before rendering
-    lda.w #command_length * 4 * 5
-    sta 0x00
-
-_clear_buffer_loop:
-    lda.w #0x00ff
-    sta.l 0x7e0000 + command_length * 4 * 5, x
-    dex
-    dex
-    dec 0
-    bne _clear_buffer_loop
-
-    sep #0x20
-
-    ldx #0x74fd  ; text buffer
-    stx 0xef50
-    lda.b #command_length  ; draw text line length used for newline
-    sta 0xef54
-
-    jsr.l messages_vwf.init_commands_list
-    jsr.l _draw_text_battle_far
-    rts
-}
+.alloc battle_commands_reloc_block in bank20_reloc {
+    .if BATTLE_CMD_VWF {
+        command_length = 6
+    } else {
+        command_length = 10
+    }
 
-_draw_single_command:
-{
-    tdc
-    sep #0x20
-    lda.w 0x0001, x  ; 0x3303
-
-    cmp #0xff
-    bne _continue
-    bra _exit
-_continue:
-    asl
-    tax
-    tdc
-    lda.l assets_battle_commands_nul_ptr, x
-    tax
-
-    lda #0x00  ; white text
-    and #0x80
-    beq _active_command
-    lda #0x04  ; gray text
-
-_active_command:
-    sta.w 0x0001, y
-    lda #0x0e  ; change tile flags
-    sta.w 0x0000, y
-    iny
-    iny
-
-_battle_command_loop:
+    draw_command_list_for_character:
+    """
+    Public RTL entry: render the per-character battle command list for the active character.
+    Dirty-bit gating moved up to the bank-02 thunk
+    (`_draw_battle_command_window_relocated`) which clears CMD_DIRTY_BIT
+    before calling here  ; gating twice would leave the inner buffer
+    filled with $00FF blanks. Inventory-open short-circuit stays.
+    """
+    ; Skip command rendering if inventory is active (bit 2 of $4A)
+    ; This prevents format buffer ($74FD) conflicts with inventory code
+        lda.l 0x7E004A
+        and #0x04
+        bne _skip_commands
+
+        lda 0x1822  ; selected character slot
+        sta 0x1816
+        phx
+        jsr.w _draw_command_list_for_character
+        plx
+
+    _skip_commands:
+        rtl
+
+    _draw_command_list_for_character:
     {
-_loop:
-    lda.l assets_battle_commands_nul_dat, x
-    cmp #0
-    beq _exit_command_loop
-    sta.w 0x0000, y
-
-    inx
-    iny
-    bra _loop
+        stz.w 0x1817
+        ldy #0x74FD  ; keep the pointer in Y this will allow to run _draw_single_command multiple times
+    ; to build {tile_flag}cmd1\n{tile_flag}cmd2\n{tile_flag}cmd3\n{tile_flag}cmd4\n{tile_flag}cmd5\0
+    ; to issue a single draw text call to the command list region.
+        lda 0x1817  ; battle command slot
+        sta 0x26
+        lda.b #command_length * 2
+        sta 0x28
+        jsr.l mult8_far  ; command_id * (command_length * 2)
+
+
+        lda 0x1817  ; slot ID
+        asl
+        asl
+        tax
+        stx 0x00
+
+    ;lda     0x1816       ; character id
+    ;asl
+    ;tax
+
+    ;rep #0x20
+    ;ldx.w #command_buffer_ptr - 10
+    ;lda.l     0x16fead,x ;cmd_text_buf_ptrs,x
+    ;stx     0xef52 ; destination ptr ?
+
+    _loop:
+        tdc
+        lda 0x1817  ; slot ID
+        asl
+        asl
+        tax
+        stx 0x00
+
+        lda 0x1816  ; character id
+
+        rep #0x20
+        asl
+        tax
+        lda.l 0x16feb7, x
+        clc
+        adc.b 0x00
+        tax
+        sep #0x20
+
+        jsr.w _draw_single_command
+        inc 0x1817
+        lda 0x1817
+        cmp #5
+        beq _exit
+        lda #1
+        sta.w 0x0000, y
+        iny
+        bra _loop
+    _exit:
+
+        tdc
+        sta.w 0x0000, y
+
+        sta 0xef55
+
+        ldx.w #command_buffer_ptr
+        stx 0xef52  ; destination
+
+        rep #0x20
+
+    ; Fill the window with 0x00ff before rendering
+        lda.w #command_length * 4 * 5
+        sta 0x00
+
+    _clear_buffer_loop:
+        lda.w #0x00ff
+        sta.l 0x7e0000 + command_length * 4 * 5, x
+        dex
+        dex
+        dec 0
+        bne _clear_buffer_loop
+
+        sep #0x20
+
+        ldx #0x74fd  ; text buffer
+        stx 0xef50
+        lda.b #command_length  ; draw text line length used for newline
+        sta 0xef54
+
+        jsr.l messages_vwf.init_commands_list
+        jsr.l _draw_text_battle_far
+        rts
     }
 
-_exit_command_loop:
-_pad_loop:
-
-
-_exit:
-    rts
+    _draw_single_command:
+    {
+        tdc
+        sep #0x20
+        lda.w 0x0001, x  ; 0x3303
+
+        cmp #0xff
+        bne _continue
+        bra _exit
+    _continue:
+        asl
+        tax
+        tdc
+        lda.l assets_battle_commands_nul_ptr, x
+        tax
+
+        lda #0x00  ; white text
+        and #0x80
+        beq _active_command
+        lda #0x04  ; gray text
+
+    _active_command:
+        sta.w 0x0001, y
+        lda #0x0e  ; change tile flags
+        sta.w 0x0000, y
+        iny
+        iny
+
+    _battle_command_loop:
+        {
+    _loop:
+        lda.l assets_battle_commands_nul_dat, x
+        cmp #0
+        beq _exit_command_loop
+        sta.w 0x0000, y
+
+        inx
+        iny
+        bra _loop
+        }
+
+    _exit_command_loop:
+    _pad_loop:
+
+
+    _exit:
+        rts
+    }
 }
diff --git a/src/battle/equip_window.s b/src/battle/equip_window.s
index 36f3ced..d97b403 100644
--- a/src/battle/equip_window.s
+++ b/src/battle/equip_window.s
@@ -1,4 +1,6 @@
 """
+.include "../bank20.i"
+
 Relocated battle equipped-items window transfer routine.
 
 Pivots the per-character equip popup from `R-label  L-label / R-item  L-item` (two columns of stacked
@@ -23,102 +25,105 @@ Tilemap dest bases (set by hooks in items_patches.s):
 
 .extern load_menu_tfr_data_trampoline
 
-tfr_equip_window_new:
-"""
-Relocated replacement for vanilla `TfrEquipWindow` at $02:97A6.
-
-Walks 6 per-block transfer passes: slot-1 / slot-2 labels into the
-left column (rows 1 and 3), slot-1 / slot-2 item top+bot halves into
-the right column (rows 0-1 and 2-3 respectively). Caller is the
-bank-02 trampoline that JSLs here  ; return via RTL.
-"""
-
 
-{
-    rep #0x10  ; X/Y 16-bit at runtime (caller may have X=1)
-    ldx.w #0xD1E8
-    stx 0x00
-    ldx.w #0xD204
-    stx 0x02
-    lda 0x1822
-    asl
-    tax
-    rep #0x20
-    lda.l 0x16FEC1, x  ; EquipTextBufPtrs
-    pha
-    lda.l 0x16FF25, x  ; RLHandTextBufPtrs
-    tax
-    sep #0x20
-
-; --- Pass A: slot-1 label -> LEFT col row 1 (buf[0..$13]) ---
-    ldy.w #0x0040
-loop_a:
-    lda.w 0x0000, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #0x0054
-    bne loop_a
-
-; --- Pass B: slot-2 label -> LEFT col row 3 (buf[$14..$27]) ---
-    ldy.w #0x00C0
-loop_b:
-    lda.w 0x0000, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #0x00D4
-    bne loop_b
-
-; --- Switch to EquipBuf ---
-    plx
-
-; --- Pass C: slot-1 item top -> RIGHT col row 0 (buf[0..$1D]) ---
-    ldy.w #0x0000
-loop_c:
-    lda.w 0x0000, x
-    sta (0x02), y
-    inx
-    iny
-    cpy.w #0x001E
-    bne loop_c
-
-; --- Pass D: slot-1 item bot -> RIGHT col row 1 (buf[$1E..$3B]) ---
-    ldy.w #0x0040
-loop_d:
-    lda.w 0x0000, x
-    sta (0x02), y
-    inx
-    iny
-    cpy.w #0x005E
-    bne loop_d
-
-; --- Pass E: slot-2 item top -> RIGHT col row 2 (buf[$3C..$59]) ---
-    ldy.w #0x0080
-loop_e:
-    lda.w 0x0000, x
-    sta (0x02), y
-    inx
-    iny
-    cpy.w #0x009E
-    bne loop_e
-
-; --- Pass F: slot-2 item bot -> RIGHT col row 3 (buf[$5A..$77]) ---
-    ldy.w #0x00C0
-loop_f:
-    lda.w 0x0000, x
-    sta (0x02), y
-    inx
-    iny
-    cpy.w #0x00DE
-    bne loop_f
-
-; --- Trigger tilemap upload (window 7, 2 transfers) ---
-    lda #0x07
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-    lda #0x01
-    sta 0x1825
-    sta 0x1824
-    rtl
+.alloc battle_equip_window_block in bank20_reloc {
+    tfr_equip_window_new:
+    """
+    Relocated replacement for vanilla `TfrEquipWindow` at $02:97A6.
+
+    Walks 6 per-block transfer passes: slot-1 / slot-2 labels into the
+    left column (rows 1 and 3), slot-1 / slot-2 item top+bot halves into
+    the right column (rows 0-1 and 2-3 respectively). Caller is the
+    bank-02 trampoline that JSLs here  ; return via RTL.
+    """
+
+
+    {
+        rep #0x10  ; X/Y 16-bit at runtime (caller may have X=1)
+        ldx.w #0xD1E8
+        stx 0x00
+        ldx.w #0xD204
+        stx 0x02
+        lda 0x1822
+        asl
+        tax
+        rep #0x20
+        lda.l 0x16FEC1, x  ; EquipTextBufPtrs
+        pha
+        lda.l 0x16FF25, x  ; RLHandTextBufPtrs
+        tax
+        sep #0x20
+
+    ; --- Pass A: slot-1 label -> LEFT col row 1 (buf[0..$13]) ---
+        ldy.w #0x0040
+    loop_a:
+        lda.w 0x0000, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #0x0054
+        bne loop_a
+
+    ; --- Pass B: slot-2 label -> LEFT col row 3 (buf[$14..$27]) ---
+        ldy.w #0x00C0
+    loop_b:
+        lda.w 0x0000, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #0x00D4
+        bne loop_b
+
+    ; --- Switch to EquipBuf ---
+        plx
+
+    ; --- Pass C: slot-1 item top -> RIGHT col row 0 (buf[0..$1D]) ---
+        ldy.w #0x0000
+    loop_c:
+        lda.w 0x0000, x
+        sta (0x02), y
+        inx
+        iny
+        cpy.w #0x001E
+        bne loop_c
+
+    ; --- Pass D: slot-1 item bot -> RIGHT col row 1 (buf[$1E..$3B]) ---
+        ldy.w #0x0040
+    loop_d:
+        lda.w 0x0000, x
+        sta (0x02), y
+        inx
+        iny
+        cpy.w #0x005E
+        bne loop_d
+
+    ; --- Pass E: slot-2 item top -> RIGHT col row 2 (buf[$3C..$59]) ---
+        ldy.w #0x0080
+    loop_e:
+        lda.w 0x0000, x
+        sta (0x02), y
+        inx
+        iny
+        cpy.w #0x009E
+        bne loop_e
+
+    ; --- Pass F: slot-2 item bot -> RIGHT col row 3 (buf[$5A..$77]) ---
+        ldy.w #0x00C0
+    loop_f:
+        lda.w 0x0000, x
+        sta (0x02), y
+        inx
+        iny
+        cpy.w #0x00DE
+        bne loop_f
+
+    ; --- Trigger tilemap upload (window 7, 2 transfers) ---
+        lda #0x07
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+        lda #0x01
+        sta 0x1825
+        sta 0x1824
+        rtl
+    }
 }
diff --git a/src/battle/graphics.s b/src/battle/graphics.s
index 11de48b..702e88b 100644
--- a/src/battle/graphics.s
+++ b/src/battle/graphics.s
@@ -1,14 +1,19 @@
 """
+.include "../bank20.i"
+
 Hard-coded battle command graphics: tile-id rows used when the row/defend command labels need to be redrawn
 from ROM.
 """
-.scope defend_row {
-    """Two glyph rows for the Defend / Row battle command labels."""
-    defend_row_length = 6
-defend_text:
-    .db 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
-    .db 222, 224, 231, 225, 226, 255
-row_text:
-    .db 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
-    .db 223, 224, 229, 227, 255, 255
+
+.alloc battle_graphics_block in bank20_reloc {
+    .scope defend_row {
+        """Two glyph rows for the Defend / Row battle command labels."""
+        defend_row_length = 6
+    defend_text:
+        .db 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+        .db 222, 224, 231, 225, 226, 255
+    row_text:
+        .db 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
+        .db 223, 224, 229, 227, 255, 255
+    }
 }
diff --git a/src/battle/inventory_budget.i b/src/battle/inventory_budget.i
new file mode 100644
index 0000000..748881e
--- /dev/null
+++ b/src/battle/inventory_budget.i
@@ -0,0 +1,21 @@
+"""
+Shared inventory VWF tile-budget constants. Included by both
+`src/battle/inventory_rolling.s` (slot init / clear loops / allocator
+clamp) and `src/battle/message.s` (NMI partial-DMA path tables).
+
+The runtime budget is validated against the items XML assets by
+`tests/test_vwf_asset_widths.py`  ; bumping ITEM_VWF_TILE_BUDGET also
+requires updating the per-slot tile_id tables to stay disjoint.
+"""
+
+
+; CHR tile_ids reserved per inventory slot (each tile_id = 16 CHR bytes).
+ITEM_VWF_TILE_BUDGET := 10
+
+; First slot's tile_id low byte. Slot N occupies
+; ITEM_VWF_TILE_BASE + N * ITEM_VWF_TILE_BUDGET .. + budget-1.
+ITEM_VWF_TILE_BASE := 0xC0
+
+; Derived sizes.
+ITEM_VWF_CHR_BYTES := ITEM_VWF_TILE_BUDGET * 16  ; CHR slice per slot
+ITEM_VWF_CHR_WORDS := ITEM_VWF_TILE_BUDGET * 8  ; word-stride clear-loop count
diff --git a/src/battle/inventory_rolling.s b/src/battle/inventory_rolling.s
index 30eab21..f46d73f 100644
--- a/src/battle/inventory_rolling.s
+++ b/src/battle/inventory_rolling.s
@@ -4,152 +4,181 @@ Battle inventory rolling-buffer engine (single column, 5 visible rows + 1 prefet
 runs the field-menu NMI DMA check.
 """
 .extern assets_items_dat
+.extern assets_items_unleashed_dat
 .extern mult8_trampoline
 .extern load_menu_tfr_data_trampoline
 .extern hex_to_dec_trampoline
 .extern normalize_num_trampoline
 .extern draw_text_rolling_trampoline
 .extern return_to_bank02
+.extern render.flush_chr_to_vram
 
-; ============================================================================
-; Rolling Buffer Implementation for Battle Inventory (Single Column)
-; ============================================================================
-;
-; Single-column layout with circular buffer, matching FF6's approach.
-; Only 5 visible rows are rendered, with edge rows pre-rendered on scroll.
-;
-; Key insight: Render BEFORE scroll animation starts (not after).
-;
-; See docs/ff6_rolling_inventory_analysis.md for design rationale.
-;
-; ============================================================================
-
-; ============================================================================
-; CONSTANTS
-; ============================================================================
-
-; Layout (single column)
-VISIBLE_ROWS := 5  ; Rows visible on screen
-BUFFER_SLOTS := 6  ; 6 slots for 5 visible (1 off-screen for pre-render)
-TOTAL_ITEMS := 48  ; Total inventory items
-TOTAL_ROWS := 48  ; One item per row now
-
-; Buffer sizes
-TEXT_BYTES_PER_ITEM := 60  ; 30 tiles x 2 bytes (dakuten + main rows)
-TILEMAP_BYTES_PER_ROW := 128  ; 2 tilemap rows x 64 bytes ($80)
-
-; Memory addresses - using freed spell list buffers
-; Spell list buffers freed by magic direct rendering: $97A6, $9E66, $A526, $ABE6, $B2A6
-text_buffer_base := 0x97A6  ; Ring buffer (6 slots × 60 = 360 bytes, uses freed spell buffer 1)
-inv_format_buffer := 0x9E66  ; Format buffer for draw_text (uses freed spell buffer 2)
-tilemap_buffer_base := 0xC4E6  ; Tilemap buffer
-tilemap_content_offset := 0x44  ; Offset to content area ($C52A - $C4E6)
-
-; ============================================================================
-; RAM VARIABLES (Using unused battle RAM)
-; ============================================================================
-; $EF97-$EF99 are explicitly marked "unused" in RAM map
-; $EF82 is marked "-" (unused)
-
-rolling_top_row := 0xEF97  ; Top visible row index (0-43)
-rolling_buffer_pos := 0xEF98  ; Circular buffer position (0-4)
-rolling_edge_row := 0xEF99  ; Row index to render (0-47)
-rolling_slot_index := 0xEF82  ; Current slot index for rendering (0-5)
-; NOTE: Using $EF82 instead of $1817 to avoid conflicts with NMI/battle commands
-
-; ============================================================================
-; VRAM CIRCULAR BUFFER SLOTS (FF6-style)
-; ============================================================================
-; 5 fixed slots for visible rows, each 128 bytes (2 tilemap rows)
-; Total VRAM usage: 640 bytes (vs 3072+ for linear)
-;
-; The HDMA scroll offset rotates which slot appears at which screen row.
-; When scrolling down: slot[circular_pos] becomes new bottom, circular_pos++
-; When scrolling up: circular_pos--, slot[circular_pos] becomes new top
-
-VRAM_SLOT_SIZE := 0x80  ; 128 bytes per slot (2 tilemap rows)
-VRAM_SLOT_BASE := 0x7400  ; Base VRAM address for slots
-
-; VRAM slot addresses - MUST match actual content positions after transfer!
-; VRAM uses WORD addresses (2 bytes per word).
-; Content goes to WRAM tilemap_buffer_base + tilemap_content_offset ($C4E6 + $44 = $C52A)
-; Entry 3 transfers from $C4E6 to VRAM $7400, size $0400 bytes = $200 words
-; WRAM byte offset $44 = VRAM word offset $22 (divide by 2)
-; So $C52A maps to VRAM $7400 + $22 = $7422
-; Each WRAM slot is $80 bytes = $40 VRAM words (64 words = 2 tilemap rows = 16 pixels)
-
-_vram_slot_table:
-    .dw 0x7422  ; Slot 0 (content at $7400 + $22 words)
-    .dw 0x7462  ; Slot 1 (+$40 words)
-    .dw 0x74A2  ; Slot 2 (+$40 words)
-    .dw 0x74E2  ; Slot 3 (+$40 words)
-    .dw 0x7522  ; Slot 4 (+$40 words)
-    .dw 0x7562  ; Slot 5 (off-screen pre-render slot)
-
-; Tilemap buffer slots (mirror of VRAM slots in WRAM)
-; Each slot is 128 bytes at tilemap_buffer_base + slot * 128
-TILEMAP_SLOT_BASE := tilemap_buffer_base + tilemap_content_offset
-
-; ============================================================================
-; init_inventory_text_buf_rolling
-; ============================================================================
-; Replacement for $029E9C - renders only 5 visible rows
-; Called when inventory window opens
-;
-; Input: None (reads $EF71 for current scroll position)
-; Output: Text buffer populated with 5 visible items
-; Clobbers: A, X, Y, $00-$06, $26-$2A
-
-init_inventory_text_buf_rolling:
-"""
-Replacement for original `$029E9C`: render the 5 visible inventory rows + 1 off-screen prefetch slot when the
-in-battle inventory window opens, resetting the rolling-buffer state.
-"""
-; Set data bank to $7E for WRAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Initialize rolling buffer state
-; At init time, always start at row 0 (top of inventory)
-; NOTE: EF65 and EF67 are initialized by ResetListScrollHDMA when inventory opens
-    stz.w 0xEF71  ; Game's row index = 0
-    stz.w 0xEF86  ; Scroll offset = 0 (top item visible)
-    stz.b 0x60  ; Cursor row = 0 (top visible row)
-    stz.w rolling_top_row
-    stz.w rolling_buffer_pos
-    ; Note: Game's $4A flag (bit 2) already indicates inventory is active
-
-; Render 6 rows to circular slots (5 visible + 1 off-screen)
-; At init, item index = slot index (both 0-5)
+.alloc battle_inventory_rolling_block in bank20_reloc {
+    .if BATTLE_ITEMS_VWF {
+        .extern messages_vwf.init_inventory
+        .extern messages_vwf.deinit
+    }
+
+    ; ============================================================================
+    ; Rolling Buffer Implementation for Battle Inventory (Single Column)
+    ; ============================================================================
+    ;
+    ; Single-column layout with circular buffer, matching FF6's approach.
+    ; Only 5 visible rows are rendered, with edge rows pre-rendered on scroll.
+    ;
+    ; Key insight: Render BEFORE scroll animation starts (not after).
+    ;
+    ; See docs/ff6_rolling_inventory_analysis.md for design rationale.
+    ;
+    ; ============================================================================
+
+    ; ============================================================================
+    ; CONSTANTS
+    ; ============================================================================
+
+    ; Layout (single column)
+    VISIBLE_ROWS := 5  ; Rows visible on screen
+    BUFFER_SLOTS := 6  ; 6 slots for 5 visible (1 off-screen for pre-render)
+    TOTAL_ITEMS := 48  ; Total inventory items
+    .include "src/battle/inventory_budget.i"
+.include "../bank20.i"
+    TOTAL_ROWS := 48  ; One item per row now
+
+    ; Buffer sizes
+    TEXT_BYTES_PER_ITEM := 60  ; 30 tiles x 2 bytes (dakuten + main rows)
+    TILEMAP_BYTES_PER_ROW := 128  ; 2 tilemap rows x 64 bytes ($80)
+
+    ; Memory addresses - using freed spell list buffers
+    ; Spell list buffers freed by magic direct rendering: $97A6, $9E66, $A526, $ABE6, $B2A6
+    text_buffer_base := 0x97A6  ; Ring buffer (6 slots × 60 = 360 bytes, uses freed spell buffer 1)
+    inv_format_buffer := 0x9E66  ; Format buffer for draw_text (uses freed spell buffer 2)
+    tilemap_buffer_base := 0xC4E6  ; Tilemap buffer
+    tilemap_content_offset := 0x46  ; Was $44 ($C52A, col 2). +2 = 1 tiles right ($C52E, col 2).
+
+    ; ============================================================================
+    ; RAM VARIABLES (Using unused battle RAM)
+    ; ============================================================================
+    ; $EF97-$EF99 are explicitly marked "unused" in RAM map
+    ; $EF82 is marked "-" (unused)
+
+    rolling_top_row := 0xEF97  ; Top visible row index (0-43)
+    rolling_buffer_pos := 0xEF98  ; Circular buffer position (0-4)
+    rolling_edge_row := 0xEF99  ; Row index to render (0-47)
+    rolling_slot_index := 0xEF82  ; Current slot index for rendering (0-5)
+    inventory_needs_full_refresh := 0xEF9D  ; Non-zero = re-render all 5 visible slots this frame.
+    ; Set on init / item swap. Scroll edges render the
+    ; new hidden slot directly via _render_*_edge_row and
+    ; do NOT need a full refresh.
+    ; NOTE: Using $EF82 instead of $1817 to avoid conflicts with NMI/battle commands
+
+    ; ============================================================================
+    ; VRAM CIRCULAR BUFFER SLOTS (FF6-style)
+    ; ============================================================================
+    ; 5 fixed slots for visible rows, each 128 bytes (2 tilemap rows)
+    ; Total VRAM usage: 640 bytes (vs 3072+ for linear)
+    ;
+    ; The HDMA scroll offset rotates which slot appears at which screen row.
+    ; When scrolling down: slot[circular_pos] becomes new bottom, circular_pos++
+    ; When scrolling up: circular_pos--, slot[circular_pos] becomes new top
+
+    VRAM_SLOT_SIZE := 0x80  ; 128 bytes per slot (2 tilemap rows)
+    VRAM_SLOT_BASE := 0x7400  ; Base VRAM address for slots
+
+    ; VRAM slot addresses - MUST match actual content positions after transfer!
+    ; VRAM uses WORD addresses (2 bytes per word).
+    ; Content goes to WRAM tilemap_buffer_base + tilemap_content_offset ($C4E6 + $44 = $C52A)
+    ; Entry 3 transfers from $C4E6 to VRAM $7400, size $0400 bytes = $200 words
+    ; WRAM byte offset $44 = VRAM word offset $22 (divide by 2)
+    ; So $C52A maps to VRAM $7400 + $22 = $7422
+    ; Each WRAM slot is $80 bytes = $40 VRAM words (64 words = 2 tilemap rows = 16 pixels)
+
+    _vram_slot_table:
+        .dw 0x7422  ; Slot 0 (content at $7400 + $22 words)
+        .dw 0x7462  ; Slot 1 (+$40 words)
+        .dw 0x74A2  ; Slot 2 (+$40 words)
+        .dw 0x74E2  ; Slot 3 (+$40 words)
+        .dw 0x7522  ; Slot 4 (+$40 words)
+        .dw 0x7562  ; Slot 5 (off-screen pre-render slot)
+
+    ; Tilemap buffer slots (mirror of VRAM slots in WRAM)
+    ; Each slot is 128 bytes at tilemap_buffer_base + slot * 128
+    TILEMAP_SLOT_BASE := tilemap_buffer_base + tilemap_content_offset
+
+    ; ============================================================================
+    ; init_inventory_text_buf_rolling
+    ; ============================================================================
+    ; Replacement for $029E9C - renders only 5 visible rows
+    ; Called when inventory window opens
+    ;
+    ; Input: None (reads $EF71 for current scroll position)
+    ; Output: Text buffer populated with 5 visible items
+    ; Clobbers: A, X, Y, $00-$06, $26-$2A
+
+    init_inventory_text_buf_rolling:
+    """
+    Replacement for original `$029E9C`: render the 5 visible inventory rows + 1 off-screen prefetch slot when the
+    in-battle inventory window opens, resetting the rolling-buffer state.
+    """
+    ; Set data bank to $7E for WRAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Initialize rolling buffer state
+    ; At init time, always start at row 0 (top of inventory)
+    ; NOTE: EF65 and EF67 are initialized by ResetListScrollHDMA when inventory opens
+        stz.w 0xEF71  ; Game's row index = 0
+        stz.w 0xEF86  ; Scroll offset = 0 (top item visible)
+        lda #1
+        sta.w inventory_needs_full_refresh  ; First frame after open: paint all visible.
+        stz.b 0x60  ; Cursor row = 0 (top visible row)
+        stz.w rolling_top_row
+        stz.w rolling_buffer_pos
+        ; Note: Game's $4A flag (bit 2) already indicates inventory is active
+
+    .if BATTLE_ITEMS_VWF {
+    ; Reset the VWF allocator to tile_id 0xC0 once for the whole pass.
+    ; The 6-slot render loop below increments the allocator naturally so
+    ; each item owns a distinct tile range (item N at 0xC0 + N * ~9 tiles).
+        jsr.l messages_vwf.init_inventory
+    }
+
+    ; Render 6 rows to circular slots (5 visible + 1 off-screen)
+    ; At init, item index = slot index (both 0-5)
     lda #0
     sta.b 0x06  ; Slot index (0-5)
 
-_init_row_loop:
-    ; For init: item index = slot index
-    lda.b 0x06
-    sta.w rolling_edge_row  ; Item index for data lookup
-    sta.w rolling_slot_index  ; Slot index for buffer position
+    _init_row_loop:
+        ; For init: item index = slot index
+        lda.b 0x06
+        sta.w rolling_edge_row  ; Item index for data lookup
+        sta.w rolling_slot_index  ; Slot index for buffer position
 
-; Save slot index
-    lda.b 0x06
-    pha
+    ; Save slot index
+        lda.b 0x06
+        pha
 
-; Render item to circular slot
-    jsr.w _render_item_to_circular_slot
+    ; Render item to circular slot
+        jsr.w _render_item_to_circular_slot
 
-; Restore slot index
-    pla
-    sta.b 0x06
+    ; Restore slot index
+        pla
+        sta.b 0x06
+
+    ; Next slot
+        inc.b 0x06
+        lda.b 0x06
+        cmp #BUFFER_SLOTS  ; 6 slots total
+        bne _init_row_loop
 
-; Next slot
-    inc.b 0x06
-    lda.b 0x06
-    cmp #BUFFER_SLOTS  ; 6 slots total
-    bne _init_row_loop
+    .if BATTLE_ITEMS_VWF {
+    ; End of inventory rolling pass: clear the VWF battle flag and signal
+    ; DMA so the inventory tile slice flushes to VRAM on the next NMI.
+        jsr.l messages_vwf.deinit
+    }
 
-; Queue VRAM transfer for initial render
+    ; Queue VRAM transfer for initial render
     lda #0x03
     ldy.w #0x0002
     jsr.l load_menu_tfr_data_trampoline
@@ -160,280 +189,286 @@ _init_row_loop:
     plb  ; Restore data bank
     rtl
 
-; ============================================================================
-; _render_inventory_item
-; ============================================================================
-; Renders a single inventory item to text buffer
-;
-; Input: rolling_slot_index = item slot index (0-47)
-; Output: Item rendered to text buffer
-; Clobbers: A, X, Y, $00-$02, $26-$2A
-
-_render_inventory_item:
-    ; Set data bank to $7E for WRAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Calculate text buffer destination
-; text_addr = text_buffer_base + (slot x TEXT_BYTES_PER_ITEM)
-    lda.w rolling_slot_index
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #text_buffer_base
-    sta.w 0xEF52  ; draw_text output destination
-    tdc
-    sep #0x20
-
-; Set up format buffer pointer
-    ldx.w #inv_format_buffer
-    stx.w 0xEF50
-
-; Set tile count for draw_text
-    lda #15  ; 15 tiles per line
-    sta.w 0xEF54
-
-; Get item data from game inventory
-; Inventory structure: 4 bytes per slot at $321A
-;   $321A+0: flags (bit 7 = disabled)
-;   $321A+1: item ID
-;   $321A+2: quantity
-;   $321A+3: ???
-    tdc  ; Clear B before TAX
-    lda.w rolling_slot_index
-    asl
-    asl  ; x 4 bytes per item
-    tax
-
-    lda.l 0x7E321B, x  ; Item ID (WRAM)
-    sta.b 0x02  ; Save for later
-    sta.b 0x26  ; For name lookup
-    lda.l 0x7E321C, x  ; Quantity (WRAM)
-    pha  ; Save quantity
-
-; Determine palette (white=enabled, gray=disabled)
-    lda #0x00
-    sta.b 0x00  ; Palette for name
-    sta.b 0x01  ; Palette for symbol
-    lda.l 0x7E321A, x  ; Flags (WRAM)
-    and #0x80
-    beq _not_disabled
-    lda #0x04  ; Gray palette
-    sta.b 0x00
-    sta.b 0x01
-
-_not_disabled:
-
-; Calculate item name address: assets_items_dat + (id x 13)
-; Must use 16-bit math since id x 13 can exceed 255
-    rep #0x20  ; 16-bit A
-    lda.b 0x02  ; Load (will get $02-$03)
-    and.w #0x00FF  ; Mask to item ID only
-    sta.b 0x08  ; Save original ID
-    asl  ; x2
-    clc
-    adc.b 0x08  ; x3
-    asl
-    asl  ; x12
-    tax
-    sep #0x20  ; Back to 8-bit
-
-; Build format string in $74FD (protected by inventory-active skip checks)
-    ldy.w #0x0000
-
-; Format code: change tile flags
-    lda #0x0E
-    sta.w inv_format_buffer, y
-    iny
-    lda.b 0x01  ; Symbol palette
-    sta.w inv_format_buffer, y
-    iny
-
-; Format code: set tile (symbol)
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.l assets_items_dat, x  ; Item symbol
-    sta.w inv_format_buffer, y
-    iny
+    ; ============================================================================
+    ; _render_inventory_item
+    ; ============================================================================
+    ; Renders a single inventory item to text buffer
+    ;
+    ; Input: rolling_slot_index = item slot index (0-47)
+    ; Output: Item rendered to text buffer
+    ; Clobbers: A, X, Y, $00-$02, $26-$2A
+
+    _render_inventory_item:
+        ; Set data bank to $7E for WRAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Calculate text buffer destination
+    ; text_addr = text_buffer_base + (slot x TEXT_BYTES_PER_ITEM)
+        lda.w rolling_slot_index
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #text_buffer_base
+        sta.w 0xEF52  ; draw_text output destination
+        tdc
+        sep #0x20
+
+    ; Set up format buffer pointer
+        ldx.w #inv_format_buffer
+        stx.w 0xEF50
+
+    ; Set tile count for draw_text
+        lda #15  ; 15 tiles per line
+        sta.w 0xEF54
+
+    ; Get item data from game inventory
+    ; Inventory structure: 4 bytes per slot at $321A
+    ;   $321A+0: flags (bit 7 = disabled)
+    ;   $321A+1: item ID
+    ;   $321A+2: quantity
+    ;   $321A+3: ???
+        tdc  ; Clear B before TAX
+        lda.w rolling_slot_index
+        asl
+        asl  ; x 4 bytes per item
+        tax
+
+        lda.l 0x7E321B, x  ; Item ID (WRAM)
+        sta.b 0x02  ; Save for later
+        sta.b 0x26  ; For name lookup
+        lda.l 0x7E321C, x  ; Quantity (WRAM)
+        pha  ; Save quantity
+
+    ; Determine palette (white=enabled, gray=disabled)
+        lda #0x00
+        sta.b 0x00  ; Palette for name
+        sta.b 0x01  ; Palette for symbol
+        lda.l 0x7E321A, x  ; Flags (WRAM)
+        and #0x80
+        beq _not_disabled
+        lda #0x04  ; Gray palette
+        sta.b 0x00
+        sta.b 0x01
+
+    _not_disabled:
+
+    ; Calculate item name address: assets_items_dat + (id x 13)
+    ; Must use 16-bit math since id x 13 can exceed 255
+        rep #0x20  ; 16-bit A
+        lda.b 0x02  ; Load (will get $02-$03)
+        and.w #0x00FF  ; Mask to item ID only
+        sta.b 0x08  ; Save original ID
+        asl  ; x2
+        clc
+        adc.b 0x08  ; x3
+        asl
+        asl  ; x12
+        tax
+        sep #0x20  ; Back to 8-bit
+
+    ; Build format string. Starts in fixed mode ; 0x0F toggles fixed <->
+    ; VWF so only the name portion routes through the VWF blitter.
+        ldy.w #0x0000
+
+        lda #0x0E
+        sta.w inv_format_buffer, y
+        iny
+        lda.b 0x01
+        sta.w inv_format_buffer, y
+        iny
+
+        lda.l assets_items_dat, x
+        sta.w inv_format_buffer, y
+        iny
+
+    .if BATTLE_ITEMS_VWF {
+        lda #0xFE
+        sta.w inv_format_buffer, y
+        iny
+    }
 
-; Format code: change tile flags (for name)
     lda #0x0E
     sta.w inv_format_buffer, y
     iny
-    lda.b 0x00  ; Name palette
+    lda.b 0x00
     sta.w inv_format_buffer, y
     iny
 
-; Copy 11 character item name
     lda #11
-    sta.b 0x00  ; Loop counter
+    sta.b 0x00
 
-_name_copy_loop:
-    inx
-    lda.l assets_items_dat, x
-    sta.w inv_format_buffer, y
-    iny
-    dec.b 0x00
-    bne _name_copy_loop
+    _name_copy_loop:
+        inx
+        lda.l assets_items_dat, x
+        sta.w inv_format_buffer, y
+        iny
+        dec.b 0x00
+        bne _name_copy_loop
+
+    .if BATTLE_ITEMS_VWF {
+        lda #0xFE
+        sta.w inv_format_buffer, y
+        iny
+        lda #0xFC
+        sta.w inv_format_buffer, y
+        iny
+        lda #12
+        sta.w inv_format_buffer, y
+        iny
+    }
 
-; Handle quantity or empty slot
-    lda.b 0x02  ; Item ID
+    lda.b 0x02
     bne _has_item
 
-; Empty slot - just finish with line break
-    pla  ; Discard quantity
-    lda #0x05  ; Line break
+    pla
+    lda #0x05
     sta.w inv_format_buffer, y
     iny
     lda #0x03
     bra _finish_format
 
-_has_item:
-    ; Add colon and quantity
-    lda #0xC8  ; Colon character ":"
-    sta.w inv_format_buffer, y
-    iny
-
-    pla  ; Get quantity
-    tax
-    jsr.l hex_to_dec_trampoline  ; Convert to decimal
-    jsr.l normalize_num_trampoline  ; Format digits
+    _has_item:
+        lda #0xC8
+        sta.w inv_format_buffer, y
+        iny
 
-; Add tens digit
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.w 0x180E  ; Tens digit
-    sta.w inv_format_buffer, y
-    iny
-
-; Add ones digit
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.w 0x180F  ; Ones digit
-
-_finish_format:
-    sta.w inv_format_buffer, y
-    iny
+        pla
+        tax
+        jsr.l hex_to_dec_trampoline
+        jsr.l normalize_num_trampoline
 
-; Terminator
-    lda #0x00
-    sta.w inv_format_buffer, y
+        lda.w 0x180E
+        sta.w inv_format_buffer, y
+        iny
 
-; Call draw_text to render formatted text to text buffer
-    jsr.l draw_text_rolling_trampoline
+        lda.w 0x180F
 
-    plb  ; Restore data bank
-    rts
-
-; ============================================================================
-; _render_inventory_item_circular
-; ============================================================================
-; Renders item to circular buffer slot
-;
-; Input: rolling_edge_row = item index (0-47) for data lookup
-;        rolling_slot_index = buffer slot (0-4) for destination
-; Output: Item rendered to text buffer at slot position
-; Clobbers: A, X, Y, $00-$02, $26-$2A
-
-_render_inventory_item_circular:
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Calculate text buffer destination using SLOT (not item index)
-; text_addr = text_buffer_base + (slot x TEXT_BYTES_PER_ITEM)
-    lda.w rolling_slot_index  ; Buffer slot (0-4)
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #text_buffer_base
-    sta.w 0xEF52  ; draw_text output destination
-    tdc
-    sep #0x20
-
-; Set up format buffer pointer
-    ldx.w #inv_format_buffer
-    stx.w 0xEF50
-
-; Set tile count for draw_text
-    lda #15  ; 15 tiles per line
-    sta.w 0xEF54
-
-; Get item data using ITEM INDEX (rolling_edge_row), not slot
-    tdc
-    lda.w rolling_edge_row  ; Item index (0-47)
-    asl
-    asl  ; x 4 bytes per item
-    tax
-
-    lda.l 0x7E321B, x  ; Item ID
-    sta.b 0x02
-    sta.b 0x26
-    lda.l 0x7E321C, x  ; Quantity
-    pha
-
-; Determine palette
-    lda #0x00
-    sta.b 0x00
-    sta.b 0x01
-    lda.l 0x7E321A, x  ; Flags
-    and #0x80
-    beq _circ_not_disabled
-    lda #0x04
-    sta.b 0x00
-    sta.b 0x01
+    _finish_format:
+        sta.w inv_format_buffer, y
+        iny
 
-_circ_not_disabled:
+        lda #0x00
+        sta.w inv_format_buffer, y
 
-; Calculate item name address: assets_items_dat + (id x 12)
-    rep #0x20
-    lda.b 0x02
-    and.w #0x00FF
-    sta.b 0x08
-    asl
-    clc
-    adc.b 0x08
-    asl
-    asl
-    tax
-    sep #0x20
-
-; Build format string in $74FD (protected by inventory-active skip checks)
-    ldy.w #0x0000
-
-; Tile flags for symbol
-    lda #0x0E
-    sta.w inv_format_buffer, y
-    iny
-    lda.b 0x01
-    sta.w inv_format_buffer, y
-    iny
+    ; Call draw_text to render formatted text to text buffer
+        jsr.l draw_text_rolling_trampoline
 
-; Symbol tile
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.l assets_items_dat, x
-    sta.w inv_format_buffer, y
-    iny
+        plb  ; Restore data bank
+        rts
 
-; Tile flags for name
+    ; ============================================================================
+    ; _render_inventory_item_circular
+    ; ============================================================================
+    ; Renders item to circular buffer slot
+    ;
+    ; Input: rolling_edge_row = item index (0-47) for data lookup
+    ;        rolling_slot_index = buffer slot (0-4) for destination
+    ; Output: Item rendered to text buffer at slot position
+    ; Clobbers: A, X, Y, $00-$02, $26-$2A
+
+    _render_inventory_item_circular:
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Calculate text buffer destination using SLOT (not item index)
+    ; text_addr = text_buffer_base + (slot x TEXT_BYTES_PER_ITEM)
+        lda.w rolling_slot_index  ; Buffer slot (0-4)
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #text_buffer_base
+        sta.w 0xEF52  ; draw_text output destination
+        tdc
+        sep #0x20
+
+    ; Set up format buffer pointer
+        ldx.w #inv_format_buffer
+        stx.w 0xEF50
+
+    ; Set tile count for draw_text
+        lda #15  ; 15 tiles per line
+        sta.w 0xEF54
+
+    ; Get item data using ITEM INDEX (rolling_edge_row), not slot
+        tdc
+        lda.w rolling_edge_row  ; Item index (0-47)
+        asl
+        asl  ; x 4 bytes per item
+        tax
+
+        lda.l 0x7E321B, x  ; Item ID
+        sta.b 0x02
+        sta.b 0x26
+        lda.l 0x7E321C, x  ; Quantity
+        pha
+
+    ; Determine palette
+        lda #0x00
+        sta.b 0x00
+        sta.b 0x01
+        lda.l 0x7E321A, x  ; Flags
+        and #0x80
+        beq _circ_not_disabled
+        lda #0x04
+        sta.b 0x00
+        sta.b 0x01
+
+    _circ_not_disabled:
+
+    ; Calculate item name address: assets_items_dat + (id x 12)
+        rep #0x20
+        lda.b 0x02
+        and.w #0x00FF
+        sta.b 0x08
+        asl
+        clc
+        adc.b 0x08
+        asl
+        asl
+        tax
+        sep #0x20
+
+    ; Build format string. The custom inventory renderer starts in fixed
+    ; mode and uses 0x0F as a fixed<->VWF toggle so only the name routes
+    ; through the VWF blitter.
+        ldy.w #0x0000
+
+    ; Tile flags for symbol (fixed mode default)
+        lda #0x0E
+        sta.w inv_format_buffer, y
+        iny
+        lda.b 0x01
+        sta.w inv_format_buffer, y
+        iny
+
+    ; Symbol tile (raw byte ; default mode is fixed)
+        lda.l assets_items_dat, x
+        sta.w inv_format_buffer, y
+        iny
+
+    ; Toggle to VWF for the name
+    .if BATTLE_ITEMS_VWF {
+        lda #0xFE
+        sta.w inv_format_buffer, y
+        iny
+    }
+
+    ; Tile flags for name
     lda #0x0E
     sta.w inv_format_buffer, y
     iny
@@ -441,19 +476,32 @@ _circ_not_disabled:
     sta.w inv_format_buffer, y
     iny
 
-; 11-char name
+    ; 11-char name (raw chars dispatched through VWF blitter)
     lda #11
     sta.b 0x00
 
-_circ_name_loop:
-    inx
-    lda.l assets_items_dat, x
-    sta.w inv_format_buffer, y
-    iny
-    dec.b 0x00
-    bne _circ_name_loop
-
-; Quantity handling
+    _circ_name_loop:
+        inx
+        lda.l assets_items_dat, x
+        sta.w inv_format_buffer, y
+        iny
+        dec.b 0x00
+        bne _circ_name_loop
+
+    ; Toggle back to fixed for colon + digits
+    .if BATTLE_ITEMS_VWF {
+        lda #0xFE
+        sta.w inv_format_buffer, y
+        iny
+        lda #0xFC
+        sta.w inv_format_buffer, y
+        iny
+        lda #12
+        sta.w inv_format_buffer, y
+        iny
+    }
+
+    ; Quantity handling
     lda.b 0x02
     bne _circ_has_item
 
@@ -464,472 +512,478 @@ _circ_name_loop:
     lda #0x03
     bra _circ_finish
 
-_circ_has_item:
-    lda #0xC8  ; Colon
-    sta.w inv_format_buffer, y
-    iny
-
-    pla
-    tax
-    jsr.l hex_to_dec_trampoline
-    jsr.l normalize_num_trampoline
-
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.w 0x180E
-    sta.w inv_format_buffer, y
-    iny
-
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.w 0x180F
-
-_circ_finish:
-    sta.w inv_format_buffer, y
-    iny
-
-    lda #0x00
-    sta.w inv_format_buffer, y
-
-    jsr.l draw_text_rolling_trampoline
-
-    plb
-    rts
-
-; ============================================================================
-; _copy_item_to_tilemap_circular
-; ============================================================================
-; Copies item from text buffer slot to tilemap slot (circular buffer)
-;
-; Input: rolling_slot_index = buffer slot (0-4)
-; Output: Item copied to tilemap buffer at slot position
-; Clobbers: A, X, Y, $00-$06
-
-_copy_item_to_tilemap_circular:
-    ; Save DBR - mult8_trampoline may change it
-    phb
-    lda #0x7E
-    pha
-    plb  ; Ensure DBR is $7E for WRAM access
-
-; Calculate tilemap buffer address for this SLOT (not item index)
-; tilemap_addr = tilemap_buffer_base + (slot x TILEMAP_BYTES_PER_ROW)
-; NO content_offset - write to full 128-byte slot
-    lda.w rolling_slot_index  ; Buffer slot (0-4)
-    sta.b 0x26
-    lda #TILEMAP_BYTES_PER_ROW
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
-    sta.b 0x00  ; Tilemap destination
-    sep #0x20
-
-; Calculate text buffer source using SLOT
-    lda.w rolling_slot_index
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    tax
-    sep #0x20
-
-; Copy first tilemap row (30 bytes)
-    ldy.w #0x0000
-
-_circ_copy_row1:
-    lda.w text_buffer_base, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #30
-    bne _circ_copy_row1
-
-; Clear remaining bytes of first row (30-63) to prevent stale border tiles
-; Fill with $FF (blank tile) and $00 (palette 0)
-
-_circ_clear_row1:
-    lda #0xFF  ; Blank tile
-    sta (0x00), y
-    iny
-    lda #0x00  ; Palette 0
-    sta (0x00), y
-    iny
-    cpy.w #0x0040
-    bne _circ_clear_row1
-
-; Copy second tilemap row (+$40 offset)
-    ldy.w #0x0040
+    _circ_has_item:
+        lda #0xC8
+        sta.w inv_format_buffer, y
+        iny
 
-_circ_copy_row2:
-    lda.w text_buffer_base, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #0x005E
-    bne _circ_copy_row2
+        pla
+        tax
+        jsr.l hex_to_dec_trampoline
+        jsr.l normalize_num_trampoline
 
-; Clear remaining bytes of second row (0x5E-0x7F)
+        lda.w 0x180E
+        sta.w inv_format_buffer, y
+        iny
 
-_circ_clear_row2:
-    lda #0xFF  ; Blank tile
-    sta (0x00), y
-    iny
-    lda #0x00  ; Palette 0
-    sta (0x00), y
-    iny
-    cpy.w #0x0080
-    bne _circ_clear_row2
-
-; Restore DBR
-    plb
-    rts
-
-; ============================================================================
-; _copy_item_to_tilemap
-; ============================================================================
-; Copies a single item from text buffer to the correct tilemap position
-; (Used for initial rendering where slot = item index)
-;
-; Input: rolling_slot_index = item slot index (0-47)
-; Output: Item copied to tilemap buffer
-; Clobbers: A, X, Y, $00-$06
-
-_copy_item_to_tilemap:
-    ; Calculate row index from slot (same as slot for single column)
-    lda.w rolling_slot_index
-    sta.b 0x04  ; Save row index
-
-; Calculate tilemap buffer address for this row
-; tilemap_addr = tilemap_buffer_base + (row x TILEMAP_BYTES_PER_ROW)
-    sta.b 0x26
-    lda #TILEMAP_BYTES_PER_ROW
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-; Add content offset (left column position)
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #tilemap_buffer_base + tilemap_content_offset
-    sta.b 0x00  ; Tilemap destination
-    sep #0x20
-
-; Calculate text buffer source offset
-    lda.w rolling_slot_index
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A  ; Get offset (slot x 60)
-    tax  ; X = offset into text buffer
-    sep #0x20
-
-; Copy first tilemap row (30 bytes = 15 tiles)
-    ldy.w #0x0000
-
-_copy_row1:
-    lda.w text_buffer_base, x  ; Read from $8EA6 + offset
-    sta (0x00), y  ; Write to tilemap buffer
-    inx
-    iny
-    cpy.w #30
-    bne _copy_row1
+        lda.w 0x180F
 
-; Copy second tilemap row (30 bytes)
-; Tilemap row 2 is at +$40 (64 bytes) from row 1
-    ldy.w #0x0040
+    _circ_finish:
+        sta.w inv_format_buffer, y
+        iny
 
-_copy_row2:
-    lda.w text_buffer_base, x  ; Read from $8EA6 + offset
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #0x005E  ; $40 + 30 = $5E
-    bne _copy_row2
-
-    rts
+        lda #0x00
+        sta.w inv_format_buffer, y
 
-; ============================================================================
-; _render_bottom_edge_row
-; ============================================================================
-; Pre-renders the new bottom row BEFORE scroll down animation starts
-; Uses circular buffer - writes to the slot that's scrolling OUT (top slot)
-; After scroll, that slot will appear at the bottom
-;
-; Called via JSL from bank 02
+        jsr.l draw_text_rolling_trampoline
 
-_render_bottom_edge_row:
-    ; Disable interrupts to prevent NMI from corrupting rolling_slot_index
-    sei
-
-; Set data bank to $7E for WRAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Calculate the new bottom item index
-; new_bottom = current_top + BUFFER_SLOTS (the item that will be at bottom after scroll)
-; With 6 slots and 5 visible, this is the item we need to pre-render
-    lda.w 0xEF71
-    clc
-    adc #BUFFER_SLOTS
-
-; Check bounds (max item is 47)
-    cmp #TOTAL_ITEMS
-    bcs _render_bottom_done  ; Past end, nothing to render
-
-    sta.w rolling_edge_row  ; Save item index for data lookup
-
-; CIRCULAR BUFFER: Write to OFF-SCREEN slot (below bottom)
-; Off-screen slot = (rolling_buffer_pos + 5) % 6
-; This slot is currently invisible and will scroll into view at bottom
-    lda.w rolling_buffer_pos
-    clc
-    adc #VISIBLE_ROWS  ; +5 to get to off-screen slot
-    cmp #BUFFER_SLOTS
-    bcc _bottom_slot_ok
-    sec
-    sbc #BUFFER_SLOTS  ; Wrap if >= 6
-
-_bottom_slot_ok:
-    sta.w rolling_slot_index  ; Slot index (0-5) for buffer positioning
-
-; Render item data to this circular slot
-    jsr.w _render_item_to_circular_slot
-
-; Queue full tilemap transfer (game's VBlank will handle it)
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-    lda #0x01
-    sta.w 0x1825
-    sta.w 0x1824
+        plb
+        rts
 
-; Advance circular position (top slot advances as we scroll down)
-    lda.w rolling_buffer_pos
-    inc
-    cmp #BUFFER_SLOTS
-    bcc _store_pos_down
-    lda #0
-
-_store_pos_down:
-    sta.w rolling_buffer_pos
-
-_render_bottom_done:
-    plb  ; Restore data bank
-    cli  ; re-enable interrupts
-    rts  ; Called from within bank $20 now
-
-; ============================================================================
-; _render_top_edge_row
-; ============================================================================
-; Pre-renders the new top row BEFORE scroll up animation starts
-; Uses circular buffer - writes to the slot that's scrolling OUT (bottom slot)
-; After scroll, that slot will appear at the top
-;
-; Called via JSL from bank 02
-
-_render_top_edge_row:
-    ; Disable interrupts to prevent NMI from corrupting rolling_slot_index
-    sei
-
-; Set data bank to $7E for WRAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Calculate the new top item index
-; new_top = current_top - 1
-    lda.w 0xEF71
-    dec  ; New top row after scroll
-
-; Check bounds
-    bmi _render_top_done  ; Negative = invalid
-
-    sta.w rolling_edge_row  ; Save item index for data lookup
-
-; CIRCULAR BUFFER: Write to slot BEFORE current circular_pos
-; This slot will become the new TOP after scroll
-; First decrement circular_pos, then write to that slot
-    lda.w rolling_buffer_pos
-    dec
-    bpl _store_pos_up
-    lda #BUFFER_SLOTS - 1  ; Wrap 0→5
-
-_store_pos_up:
-    sta.w rolling_buffer_pos  ; Update position FIRST for scroll up
-    sta.w rolling_slot_index  ; Slot index (0-5) for buffer positioning
+    ; ============================================================================
+    ; _copy_item_to_tilemap_circular
+    ; ============================================================================
+    ; Copies item from text buffer slot to tilemap slot (circular buffer)
+    ;
+    ; Input: rolling_slot_index = buffer slot (0-4)
+    ; Output: Item copied to tilemap buffer at slot position
+    ; Clobbers: A, X, Y, $00-$06
+
+    _copy_item_to_tilemap_circular:
+        ; Save DBR - mult8_trampoline may change it
+        phb
+        lda #0x7E
+        pha
+        plb  ; Ensure DBR is $7E for WRAM access
+
+    ; Calculate tilemap buffer address for this SLOT (not item index)
+    ; tilemap_addr = tilemap_buffer_base + (slot x TILEMAP_BYTES_PER_ROW)
+    ; NO content_offset - write to full 128-byte slot
+        lda.w rolling_slot_index  ; Buffer slot (0-4)
+        sta.b 0x26
+        lda #TILEMAP_BYTES_PER_ROW
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
+        sta.b 0x00  ; Tilemap destination
+        sep #0x20
+
+    ; Calculate text buffer source using SLOT
+        lda.w rolling_slot_index
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        tax
+        sep #0x20
+
+    ; Copy first tilemap row (30 bytes)
+        ldy.w #0x0000
+
+    _circ_copy_row1:
+        lda.w text_buffer_base, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #30
+        bne _circ_copy_row1
+
+    ; Clear remaining bytes of first row (30-63) to prevent stale border tiles
+    ; Fill with $FF (blank tile) and $00 (palette 0)
+
+    _circ_clear_row1:
+        lda #0xFF  ; Blank tile
+        sta (0x00), y
+        iny
+        lda #0x00  ; Palette 0
+        sta (0x00), y
+        iny
+        cpy.w #0x0040
+        bne _circ_clear_row1
+
+    ; Copy second tilemap row (+$40 offset)
+        ldy.w #0x0040
+
+    _circ_copy_row2:
+        lda.w text_buffer_base, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #0x005E
+        bne _circ_copy_row2
+
+    ; Clear remaining bytes of second row (0x5E-0x7F)
+
+    _circ_clear_row2:
+        lda #0xFF  ; Blank tile
+        sta (0x00), y
+        iny
+        lda #0x00  ; Palette 0
+        sta (0x00), y
+        iny
+        cpy.w #0x0080
+        bne _circ_clear_row2
+
+    ; Restore DBR
+        plb
+        rts
+
+    ; ============================================================================
+    ; _copy_item_to_tilemap
+    ; ============================================================================
+    ; Copies a single item from text buffer to the correct tilemap position
+    ; (Used for initial rendering where slot = item index)
+    ;
+    ; Input: rolling_slot_index = item slot index (0-47)
+    ; Output: Item copied to tilemap buffer
+    ; Clobbers: A, X, Y, $00-$06
+
+    _copy_item_to_tilemap:
+        ; Calculate row index from slot (same as slot for single column)
+        lda.w rolling_slot_index
+        sta.b 0x04  ; Save row index
+
+    ; Calculate tilemap buffer address for this row
+    ; tilemap_addr = tilemap_buffer_base + (row x TILEMAP_BYTES_PER_ROW)
+        sta.b 0x26
+        lda #TILEMAP_BYTES_PER_ROW
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+    ; Add content offset (left column position)
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #tilemap_buffer_base + tilemap_content_offset
+        sta.b 0x00  ; Tilemap destination
+        sep #0x20
+
+    ; Calculate text buffer source offset
+        lda.w rolling_slot_index
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A  ; Get offset (slot x 60)
+        tax  ; X = offset into text buffer
+        sep #0x20
+
+    ; Copy first tilemap row (30 bytes = 15 tiles)
+        ldy.w #0x0000
+
+    _copy_row1:
+        lda.w text_buffer_base, x  ; Read from $8EA6 + offset
+        sta (0x00), y  ; Write to tilemap buffer
+        inx
+        iny
+        cpy.w #30
+        bne _copy_row1
+
+    ; Copy second tilemap row (30 bytes)
+    ; Tilemap row 2 is at +$40 (64 bytes) from row 1
+        ldy.w #0x0040
+
+    _copy_row2:
+        lda.w text_buffer_base, x  ; Read from $8EA6 + offset
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #0x005E  ; $40 + 30 = $5E
+        bne _copy_row2
+
+        rts
+
+    ; ============================================================================
+    ; _render_bottom_edge_row
+    ; ============================================================================
+    ; Pre-renders the new bottom row BEFORE scroll down animation starts
+    ; Uses circular buffer - writes to the slot that's scrolling OUT (top slot)
+    ; After scroll, that slot will appear at the bottom
+    ;
+    ; Called via JSL from bank 02
 
-; Render item data to this circular slot
-    jsr.w _render_item_to_circular_slot
+    _render_bottom_edge_row:
+        ; Disable interrupts to prevent NMI from corrupting rolling_slot_index
+        sei
 
-; Queue full tilemap transfer (game's VBlank will handle it)
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-    lda #0x01
-    sta.w 0x1825
-    sta.w 0x1824
+    ; Set data bank to $7E for WRAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Calculate the new bottom item index
+    ; new_bottom = current_top + BUFFER_SLOTS (the item that will be at bottom after scroll)
+    ; With 6 slots and 5 visible, this is the item we need to pre-render
+        lda.w 0xEF71
+        clc
+        adc #BUFFER_SLOTS
+
+    ; Check bounds (max item is 47)
+        cmp #TOTAL_ITEMS
+        bcs _render_bottom_done  ; Past end, nothing to render
+
+        sta.w rolling_edge_row  ; Save item index for data lookup
+
+    ; CIRCULAR BUFFER: Write to OFF-SCREEN slot (below bottom)
+    ; Off-screen slot = (rolling_buffer_pos + 5) % 6
+    ; This slot is currently invisible and will scroll into view at bottom
+        lda.w rolling_buffer_pos
+        clc
+        adc #VISIBLE_ROWS  ; +5 to get to off-screen slot
+        cmp #BUFFER_SLOTS
+        bcc _bottom_slot_ok
+        sec
+        sbc #BUFFER_SLOTS  ; Wrap if >= 6
+
+    _bottom_slot_ok:
+        sta.w rolling_slot_index  ; Slot index (0-5) for buffer positioning
+
+    ; Render item data to this circular slot
+        jsr.w _render_item_to_circular_slot
+
+    ; Queue full tilemap transfer (game's VBlank will handle it)
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+        lda #0x01
+        sta.w 0x1825
+        sta.w 0x1824
+
+    ; Advance circular position (top slot advances as we scroll down)
+        lda.w rolling_buffer_pos
+        inc
+        cmp #BUFFER_SLOTS
+        bcc _store_pos_down
+        lda #0
+
+    _store_pos_down:
+        sta.w rolling_buffer_pos
+
+    _render_bottom_done:
+        plb  ; Restore data bank
+        cli  ; re-enable interrupts
+        rts  ; Called from within bank $20 now
+
+    ; ============================================================================
+    ; _render_top_edge_row
+    ; ============================================================================
+    ; Pre-renders the new top row BEFORE scroll up animation starts
+    ; Uses circular buffer - writes to the slot that's scrolling OUT (bottom slot)
+    ; After scroll, that slot will appear at the top
+    ;
+    ; Called via JSL from bank 02
 
-_render_top_done:
-    plb  ; Restore data bank
-    cli  ; re-enable interrupts
-    rts  ; Called from within bank $20 now
-
-; ============================================================================
-; _render_item_to_circular_slot
-; ============================================================================
-; Renders item to a circular buffer slot
-;
-; Input: rolling_edge_row = item index (0-47) for data lookup
-;        rolling_slot_index = slot index (0-4) for destination
-; Output: Item rendered to text buffer slot, copied to tilemap slot
-
-_render_item_to_circular_slot:
-    ; Save processor status and disable interrupts
-    ; Using PHP/PLP instead of SEI/CLI to handle nested calls correctly
-    php
-    sei
-
-; Save DBR - draw_text may change it and we need it for tilemap copy
-    phb
-
-; Save D and set to $0000 for direct page operations
-    rep #0x20
-    tdc
-    pha  ; Save original D on stack
-    lda.w #0x0000
-    tcd
-    sep #0x20
-
-; CRITICAL: Save zero page variables on stack
-; This prevents corruption if NMI fires during game's block copy
-; Save $00-$0B (block copy uses $00-$06) and $26-$2B (Mult8 uses these)
-    lda.b 0x00
-    pha
-    lda.b 0x01
-    pha
-    lda.b 0x02
-    pha
-    lda.b 0x03
-    pha
-    lda.b 0x04
-    pha
-    lda.b 0x05
-    pha
-    lda.b 0x06
-    pha
-    lda.b 0x07
-    pha
-    lda.b 0x08
-    pha
-    lda.b 0x09
-    pha
-    lda.b 0x0A
-    pha
-    lda.b 0x0B
-    pha
-    ; Also save $26-$2B used by mult8_trampoline
-    lda.b 0x26
-    pha
-    lda.b 0x27
-    pha
-    lda.b 0x28
-    pha
-    lda.b 0x29
-    pha
-    lda.b 0x2A
-    pha
-    lda.b 0x2B
-    pha
-
-; Calculate text buffer destination for this SLOT
-; text_addr = text_buffer_base + (slot × TEXT_BYTES_PER_ITEM)
-    lda.w rolling_slot_index  ; Slot index (0-4)
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM  ; 60 bytes per slot
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #text_buffer_base
-    sta.w 0xEF52  ; draw_text output destination
-    tdc
-    sep #0x20
-
-; Set up format buffer pointer
-    ldx.w #inv_format_buffer
-    stx.w 0xEF50
-    lda #15
-    sta.w 0xEF54
-
-; Get item data using rolling_edge_row (the actual item index)
-    tdc
-    lda.w rolling_edge_row
-    asl
-    asl
-    tax
-
-    lda.l 0x7E321B, x  ; Item ID
-    sta.b 0x02
-    lda.l 0x7E321C, x  ; Quantity
-    pha
-
-; Palette selection
-    lda #0x00
-    sta.b 0x00
-    sta.b 0x01
-    lda.l 0x7E321A, x
-    and #0x80
-    beq _slot_not_disabled
-    lda #0x04
-    sta.b 0x00
-    sta.b 0x01
+    _render_top_edge_row:
+        ; Disable interrupts to prevent NMI from corrupting rolling_slot_index
+        sei
 
-_slot_not_disabled:
+    ; Set data bank to $7E for WRAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Calculate the new top item index
+    ; new_top = current_top - 1
+        lda.w 0xEF71
+        dec  ; New top row after scroll
+
+    ; Check bounds
+        bmi _render_top_done  ; Negative = invalid
+
+        sta.w rolling_edge_row  ; Save item index for data lookup
+
+    ; CIRCULAR BUFFER: Write to slot BEFORE current circular_pos
+    ; This slot will become the new TOP after scroll
+    ; First decrement circular_pos, then write to that slot
+        lda.w rolling_buffer_pos
+        dec
+        bpl _store_pos_up
+        lda #BUFFER_SLOTS - 1  ; Wrap 0→5
+
+    _store_pos_up:
+        sta.w rolling_buffer_pos  ; Update position FIRST for scroll up
+        sta.w rolling_slot_index  ; Slot index (0-5) for buffer positioning
+
+    ; Render item data to this circular slot
+        jsr.w _render_item_to_circular_slot
+
+    ; Queue full tilemap transfer (game's VBlank will handle it)
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+        lda #0x01
+        sta.w 0x1825
+        sta.w 0x1824
+
+    _render_top_done:
+        plb  ; Restore data bank
+        cli  ; re-enable interrupts
+        rts  ; Called from within bank $20 now
+
+    ; ============================================================================
+    ; _render_item_to_circular_slot
+    ; ============================================================================
+    ; Renders item to a circular buffer slot
+    ;
+    ; Input: rolling_edge_row = item index (0-47) for data lookup
+    ;        rolling_slot_index = slot index (0-4) for destination
+    ; Output: Item rendered to text buffer slot, copied to tilemap slot
+
+    _render_item_to_circular_slot:
+        ; Save processor status and disable interrupts
+        ; Using PHP/PLP instead of SEI/CLI to handle nested calls correctly
+        php
+        sei
+
+    ; Save DBR - draw_text may change it and we need it for tilemap copy
+        phb
+
+    ; Save D and set to $0000 for direct page operations
+        rep #0x20
+        tdc
+        pha  ; Save original D on stack
+        lda.w #0x0000
+        tcd
+        sep #0x20
+
+    ; CRITICAL: Save zero page variables on stack
+    ; This prevents corruption if NMI fires during game's block copy
+    ; Save $00-$0B (block copy uses $00-$06) and $26-$2B (Mult8 uses these)
+        lda.b 0x00
+        pha
+        lda.b 0x01
+        pha
+        lda.b 0x02
+        pha
+        lda.b 0x03
+        pha
+        lda.b 0x04
+        pha
+        lda.b 0x05
+        pha
+        lda.b 0x06
+        pha
+        lda.b 0x07
+        pha
+        lda.b 0x08
+        pha
+        lda.b 0x09
+        pha
+        lda.b 0x0A
+        pha
+        lda.b 0x0B
+        pha
+        ; Also save $26-$2B used by mult8_trampoline
+        lda.b 0x26
+        pha
+        lda.b 0x27
+        pha
+        lda.b 0x28
+        pha
+        lda.b 0x29
+        pha
+        lda.b 0x2A
+        pha
+        lda.b 0x2B
+        pha
+
+    ; Calculate text buffer destination for this SLOT
+    ; text_addr = text_buffer_base + (slot × TEXT_BYTES_PER_ITEM)
+        lda.w rolling_slot_index  ; Slot index (0-4)
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM  ; 60 bytes per slot
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #text_buffer_base
+        sta.w 0xEF52  ; draw_text output destination
+        tdc
+        sep #0x20
+
+    ; Set up format buffer pointer
+        ldx.w #inv_format_buffer
+        stx.w 0xEF50
+        lda #15
+        sta.w 0xEF54
+
+    ; Get item data using rolling_edge_row (the actual item index)
+        tdc
+        lda.w rolling_edge_row
+        asl
+        asl
+        tax
+
+        lda.l 0x7E321B, x  ; Item ID
+        sta.b 0x02
+        bne _slot_id_nonzero
+        jmp.w _empty_slot_fast  ; ID == 0 -> empty entry, skip VWF
+
+    _slot_id_nonzero:
+        lda.l 0x7E321C, x  ; Quantity
+        pha
+
+    ; Palette selection
+        lda #0x00
+        sta.b 0x00
+        sta.b 0x01
+        lda.l 0x7E321A, x
+        and #0x80
+        beq _slot_not_disabled
+        lda #0x04
+        sta.b 0x00
+        sta.b 0x01
+
+    _slot_not_disabled:
+
+    ; Calculate item name offset into the 17-byte-per-record
+    ; assets_items_unleashed_dat table: id * 17 = (id << 4) + id.
+        rep #0x20
+        lda.b 0x02
+        and.w #0x00FF
+        sta.b 0x08
+        asl
+        asl
+        asl
+        asl
+        clc
+        adc.b 0x08
+        tax
+        sep #0x20
+
+    ; Build format string. Same fixed-default + 0x0F-toggle pattern the
+    ; other two format builders in this file use.
+        ldy.w #0x0000
+        lda #0x0E
+        sta.w inv_format_buffer, y
+        iny
+        lda.b 0x01
+        sta.w inv_format_buffer, y
+        iny
+
+        lda.l assets_items_unleashed_dat, x
+        sta.w inv_format_buffer, y
+        iny
+
+    .if BATTLE_ITEMS_VWF {
+        lda #0xFE
+        sta.w inv_format_buffer, y
+        iny
+    }
 
-; Calculate item name offset
-    rep #0x20
-    lda.b 0x02
-    and.w #0x00FF
-    sta.b 0x08
-    asl
-    clc
-    adc.b 0x08
-    asl
-    asl
-    tax
-    sep #0x20
-
-; Build format string
-    ldy.w #0x0000
-    lda #0x0E
-    sta.w inv_format_buffer, y
-    iny
-    lda.b 0x01
-    sta.w inv_format_buffer, y
-    iny
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.l assets_items_dat, x
-    sta.w inv_format_buffer, y
-    iny
     lda #0x0E
     sta.w inv_format_buffer, y
     iny
@@ -937,16 +991,28 @@ _slot_not_disabled:
     sta.w inv_format_buffer, y
     iny
 
-    lda #11
+    lda #16
     sta.b 0x00
 
-_slot_name_loop:
-    inx
-    lda.l assets_items_dat, x
-    sta.w inv_format_buffer, y
-    iny
-    dec.b 0x00
-    bne _slot_name_loop
+    _slot_name_loop:
+        inx
+        lda.l assets_items_unleashed_dat, x
+        sta.w inv_format_buffer, y
+        iny
+        dec.b 0x00
+        bne _slot_name_loop
+
+    .if BATTLE_ITEMS_VWF {
+        lda #0xFE
+        sta.w inv_format_buffer, y
+        iny
+        lda #0xFC
+        sta.w inv_format_buffer, y
+        iny
+        lda #12
+        sta.w inv_format_buffer, y
+        iny
+    }
 
     lda.b 0x02
     bne _slot_has_item
@@ -957,1538 +1023,1595 @@ _slot_name_loop:
     lda #0x03
     bra _slot_finish
 
-_slot_has_item:
-    lda #0xC8
-    sta.w inv_format_buffer, y
-    iny
-    pla
-    tax
-    jsr.l hex_to_dec_trampoline
-    jsr.l normalize_num_trampoline
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.w 0x180E
-    sta.w inv_format_buffer, y
-    iny
-    lda #0x03
-    sta.w inv_format_buffer, y
-    iny
-    lda.w 0x180F
-
-_slot_finish:
-    sta.w inv_format_buffer, y
-    iny
-    lda #0x00
-    sta.w inv_format_buffer, y
-
-    jsr.l draw_text_rolling_trampoline
-
-; draw_text fills text buffer - tilemap copy is done separately by:
-; - _copy_all_slots_to_tilemap in tfr_inventory_list_rolling (for init)
-; - _copy_slot_to_tilemap in scroll hooks (for scrolling)
-
-; CRITICAL: Restore zero page variables from stack (reverse order)
-; First restore $26-$2B (last pushed)
-    pla
-    sta.b 0x2B
-    pla
-    sta.b 0x2A
-    pla
-    sta.b 0x29
-    pla
-    sta.b 0x28
-    pla
-    sta.b 0x27
-    pla
-    sta.b 0x26
-    ; Then restore $00-$0B
-    pla
-    sta.b 0x0B
-    pla
-    sta.b 0x0A
-    pla
-    sta.b 0x09
-    pla
-    sta.b 0x08
-    pla
-    sta.b 0x07
-    pla
-    sta.b 0x06
-    pla
-    sta.b 0x05
-    pla
-    sta.b 0x04
-    pla
-    sta.b 0x03
-    pla
-    sta.b 0x02
-    pla
-    sta.b 0x01
-    pla
-    sta.b 0x00
-
-; Restore D before returning
-    rep #0x20
-    pla
-    tcd
-    sep #0x20
-
-; Restore DBR (draw_text may have changed it)
-    plb
-
-; Restore processor status (including interrupt flag)
-    plp
-    rts
-
-; ============================================================================
-; _transfer_circular_slot
-; ============================================================================
-; Transfers a single circular slot to its fixed VRAM address
-; Uses direct DMA instead of menu transfer system for precise control
-;
-; Input: rolling_slot_index = slot index (0-4)
-; Output: 128 bytes transferred to VRAM
-
-_transfer_circular_slot:
-    ; Save processor status and disable interrupts
-    php
-    sei
-
-; Save DBR - mult8_trampoline may change it
-    phb
-    lda #0x7E
-    pha
-    plb  ; Ensure DBR is $7E for WRAM access
-
-; CRITICAL: Save zero page variables on stack
-; This prevents corruption if NMI fires during game's block copy
-; Save $00-$0B (block copy uses $00-$06) and $26-$2B (Mult8 uses these)
-    lda.b 0x00
-    pha
-    lda.b 0x01
-    pha
-    lda.b 0x02
-    pha
-    lda.b 0x03
-    pha
-    lda.b 0x04
-    pha
-    lda.b 0x05
-    pha
-    lda.b 0x06
-    pha
-    lda.b 0x07
-    pha
-    lda.b 0x08
-    pha
-    lda.b 0x09
-    pha
-    lda.b 0x0A
-    pha
-    lda.b 0x0B
-    pha
-    ; Also save $26-$2B used by mult8_trampoline
-    lda.b 0x26
-    pha
-    lda.b 0x27
-    pha
-    lda.b 0x28
-    pha
-    lda.b 0x29
-    pha
-    lda.b 0x2A
-    pha
-    lda.b 0x2B
-    pha
-
-; Look up VRAM destination from slot table
-    lda.w rolling_slot_index
-    asl  ; ×2 for word lookup
-    tax
-    rep #0x20
-    lda.l _vram_slot_table, x  ; Get VRAM address ($7400, $7480, etc.)
-    sta.b 0x04  ; Save VRAM dest
-    sep #0x20
-
-; Calculate tilemap buffer source
-; source = tilemap_buffer_base + (slot × 128) - NO content_offset!
-    lda.w rolling_slot_index
-    sta.b 0x26
-    lda #TILEMAP_BYTES_PER_ROW
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
-    sta.b 0x00  ; Source address (low word)
-    sep #0x20
-
-; Queue using menu transfer system (entry 3 covers our slots)
-; The transfer will include this slot since it's at the right offset
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-
-    lda #0x01
-    sta.w 0x1825
-    sta.w 0x1824
-
-; CRITICAL: Restore zero page variables from stack (reverse order)
-; First restore $26-$2B (last pushed)
-    pla
-    sta.b 0x2B
-    pla
-    sta.b 0x2A
-    pla
-    sta.b 0x29
-    pla
-    sta.b 0x28
-    pla
-    sta.b 0x27
-    pla
-    sta.b 0x26
-    ; Then restore $00-$0B
-    pla
-    sta.b 0x0B
-    pla
-    sta.b 0x0A
-    pla
-    sta.b 0x09
-    pla
-    sta.b 0x08
-    pla
-    sta.b 0x07
-    pla
-    sta.b 0x06
-    pla
-    sta.b 0x05
-    pla
-    sta.b 0x04
-    pla
-    sta.b 0x03
-    pla
-    sta.b 0x02
-    pla
-    sta.b 0x01
-    pla
-    sta.b 0x00
-
-; Restore DBR
-    plb
-
-; Restore processor status (including interrupt flag)
-    plp
-    rts
-
-; ============================================================================
-; _copy_slot_to_tilemap
-; ============================================================================
-; Copies a single slot from text buffer to tilemap buffer
-; Input: rolling_slot_index = slot index (0-5)
-; Uses $00-$01, $26, $28, $2A, X, Y
-
-_copy_slot_to_tilemap:
-    ; Save processor status and disable interrupts
-    ; Using PHP/PLP instead of SEI/CLI to handle nested calls correctly
-    php
-    sei
-
-; Save DBR - mult8_trampoline may change it
-    phb
-
-; Save and set D to 0 for direct page operations
-    rep #0x20
-    tdc
-    pha
-    lda.w #0x0000
-    tcd
-    sep #0x20
-
-; CRITICAL: Save zero page variables on stack
-; This prevents corruption if NMI fires during game's block copy
-; Save $00-$0B (block copy uses $00-$06) and $26-$2B (Mult8 uses these)
-    lda.b 0x00
-    pha
-    lda.b 0x01
-    pha
-    lda.b 0x02
-    pha
-    lda.b 0x03
-    pha
-    lda.b 0x04
-    pha
-    lda.b 0x05
-    pha
-    lda.b 0x06
-    pha
-    lda.b 0x07
-    pha
-    lda.b 0x08
-    pha
-    lda.b 0x09
-    pha
-    lda.b 0x0A
-    pha
-    lda.b 0x0B
-    pha
-    ; Also save $26-$2B used by mult8_trampoline
-    lda.b 0x26
-    pha
-    lda.b 0x27
-    pha
-    lda.b 0x28
-    pha
-    lda.b 0x29
-    pha
-    lda.b 0x2A
-    pha
-    lda.b 0x2B
-    pha
-
-; Calculate tilemap destination (NO content_offset - write to full slot)
-    lda.w rolling_slot_index
-    sta.b 0x26
-    lda #TILEMAP_BYTES_PER_ROW
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
-    sta.b 0x00
-    sep #0x20
-
-; Calculate text buffer source
-    lda.w rolling_slot_index
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    tax
-    sep #0x20
-
-; Copy row 1 (30 bytes) - content only, preserve borders
-    ldy.w #0x0000
-
-_copy_slot_row1:
-    lda.w text_buffer_base, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #30
-    bne _copy_slot_row1
-
-; Copy row 2 (+$40 offset) - content only, preserve borders
-    ldy.w #0x0040
-
-_copy_slot_row2:
-    lda.w text_buffer_base, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #0x005E
-    bne _copy_slot_row2
-
-; CRITICAL: Restore zero page variables $26-$2B from stack (reverse order - pushed last, pop first)
-    pla
-    sta.b 0x2B
-    pla
-    sta.b 0x2A
-    pla
-    sta.b 0x29
-    pla
-    sta.b 0x28
-    pla
-    sta.b 0x27
-    pla
-    sta.b 0x26
-
-; CRITICAL: Restore zero page variables $00-$0B from stack (reverse order)
-    pla
-    sta.b 0x0B
-    pla
-    sta.b 0x0A
-    pla
-    sta.b 0x09
-    pla
-    sta.b 0x08
-    pla
-    sta.b 0x07
-    pla
-    sta.b 0x06
-    pla
-    sta.b 0x05
-    pla
-    sta.b 0x04
-    pla
-    sta.b 0x03
-    pla
-    sta.b 0x02
-    pla
-    sta.b 0x01
-    pla
-    sta.b 0x00
-
-; Restore D register
-    rep #0x20
-    pla
-    tcd
-    sep #0x20
-
-; Restore DBR (pushed after php/sei)
-    plb
-
-; Restore processor status (including interrupt flag)
-    plp
-    rts
+    _slot_has_item:
+        lda #0xC8
+        sta.w inv_format_buffer, y
+        iny
+        pla
+        ; M=8 pla loads A.lo only; A.hi retains a leaked byte from the
+        ; preceding 16-bit arithmetic. With X=16 the next `tax` would copy
+        ; that garbage into X.hi and hex_to_dec would emit BCD for the
+        ; wider value (qty 1 + leaked $06 -> X=$0601 -> "37"). Zero-extend
+        ; A explicitly before transfer.
+        rep #0x20
+        and.w #0x00ff
+        tax
+        sep #0x20
+        jsr.l hex_to_dec_trampoline
+        jsr.l normalize_num_trampoline
+        lda.w 0x180E
+        sta.w inv_format_buffer, y
+        iny
+        lda.w 0x180F
+
+    _slot_finish:
+        sta.w inv_format_buffer, y
+        iny
+        lda #0x00
+        sta.w inv_format_buffer, y
+
+        jsr.l draw_text_rolling_trampoline
+        jmp.w _slot_render_done
+
+    _empty_slot_fast:
+    """
+    Empty inventory entry (item ID = 0). Skip the format-build + VWF
+    blit and fill the slot's text buffer with 30 blank tile entries
+    (2 bytes each = `$FF $00`). Saves the ~80K cycles draw_text would
+    otherwise burn rendering an empty palette + name string.
+    """
+        rep #0x20
+        lda.w 0xEF52
+        sta.b 0x04  ; ptr-to-slot in DP scratch
+        sep #0x20
+        ldy.w #0
+
+    _empty_fill_loop:
+        lda.b #0xFF
+        sta (0x04), y
+        iny
+        lda.b #0x00
+        sta (0x04), y
+        iny
+        cpy.w #60
+        bne _empty_fill_loop
+
+    _slot_render_done:
+
+    ; draw_text fills text buffer - tilemap copy is done separately by:
+    ; - _copy_all_slots_to_tilemap in tfr_inventory_list_rolling (for init)
+    ; - _copy_slot_to_tilemap in scroll hooks (for scrolling)
+
+    ; CRITICAL: Restore zero page variables from stack (reverse order)
+    ; First restore $26-$2B (last pushed)
+        pla
+        sta.b 0x2B
+        pla
+        sta.b 0x2A
+        pla
+        sta.b 0x29
+        pla
+        sta.b 0x28
+        pla
+        sta.b 0x27
+        pla
+        sta.b 0x26
+        ; Then restore $00-$0B
+        pla
+        sta.b 0x0B
+        pla
+        sta.b 0x0A
+        pla
+        sta.b 0x09
+        pla
+        sta.b 0x08
+        pla
+        sta.b 0x07
+        pla
+        sta.b 0x06
+        pla
+        sta.b 0x05
+        pla
+        sta.b 0x04
+        pla
+        sta.b 0x03
+        pla
+        sta.b 0x02
+        pla
+        sta.b 0x01
+        pla
+        sta.b 0x00
+
+    ; Restore D before returning
+        rep #0x20
+        pla
+        tcd
+        sep #0x20
+
+    ; Restore DBR (draw_text may have changed it)
+        plb
+
+    ; Restore processor status (including interrupt flag)
+        plp
+        rts
+
+    ; ============================================================================
+    ; _transfer_circular_slot
+    ; ============================================================================
+    ; Transfers a single circular slot to its fixed VRAM address
+    ; Uses direct DMA instead of menu transfer system for precise control
+    ;
+    ; Input: rolling_slot_index = slot index (0-4)
+    ; Output: 128 bytes transferred to VRAM
 
-; ============================================================================
-; _copy_all_slots_to_tilemap
-; ============================================================================
-; Copies all 6 slots from text buffer to tilemap buffer
-; Called from tfr_inventory_list_rolling AFTER game clears window buffers
+    _transfer_circular_slot:
+        ; Save processor status and disable interrupts
+        php
+        sei
 
-_copy_all_slots_to_tilemap:
     ; Save DBR - mult8_trampoline may change it
-    phb
-    lda #0x7E
-    pha
-    plb  ; Ensure DBR is $7E for WRAM access
-
-    lda #0
-    sta.b 0x06  ; Slot counter (0-5)
-
-_copy_slots_loop:
-    ; Calculate tilemap destination: tilemap_buffer_base + (slot × 128)
-    ; NO content_offset - write to full 128-byte slot
-    lda.b 0x06
-    sta.b 0x26
-    lda #TILEMAP_BYTES_PER_ROW  ; 128
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    clc
-    adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
-    sta.b 0x00  ; Tilemap dest pointer
-    sep #0x20
-
-; Calculate text buffer source: text_buffer_base + (slot × 60)
-    lda.b 0x06
-    sta.b 0x26
-    lda #TEXT_BYTES_PER_ITEM  ; 60
-    sta.b 0x28
-    jsr.l mult8_trampoline
-
-    rep #0x20
-    lda.b 0x2A
-    tax  ; X = text buffer offset
-    sep #0x20
-
-; Copy row 1 (30 bytes) - content only, preserve borders
-    ldy.w #0x0000
-
-_copy_all_row1:
-    lda.w text_buffer_base, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #30
-    bne _copy_all_row1
-
-; Copy row 2 (+$40 offset in tilemap) - content only, preserve borders
-    ldy.w #0x0040
-
-_copy_all_row2:
-    lda.w text_buffer_base, x
-    sta (0x00), y
-    inx
-    iny
-    cpy.w #0x005E
-    bne _copy_all_row2
-
-; Next slot
-    inc.b 0x06
-    lda.b 0x06
-    cmp #BUFFER_SLOTS  ; 6 slots
-    bne _copy_slots_loop
-
-; Restore DBR
-    plb
-    rts
-
-; ============================================================================
-; tfr_inventory_list_rolling
-; ============================================================================
-; Replacement for TfrInventoryList ($0298FA)
-; Transfers the visible portion of tilemap buffer to VRAM
-;
-; Called via JSL from bank 02
-
-tfr_inventory_list_rolling:
-"""
-Replacement for original `TfrInventoryList` ($0298FA): re-render the visible inventory rows into our buffer and DMA
-the slice to VRAM. Reached via JSL from bank 02.
-"""
-; Set data bank to $7E for WRAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; RE-RENDER all visible items to our rolling buffer
-; This is necessary because DrawInventoryItemText (called after item swap)
-; writes to the OLD text buffer at $8EA6, not our buffer at $97A6.
-; By always re-rendering here, swapped items display correctly.
-    jsr.w _refresh_visible_items_internal
-
-; Copy all 6 slots from text buffer to tilemap buffer
-; This runs AFTER the game's window clearing at $9AF4
-    jsr.w _copy_all_slots_to_tilemap
-
-; Queue VRAM transfer (entry 3 covers $7400-$77FF)
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
+        phb
+        lda #0x7E
+        pha
+        plb  ; Ensure DBR is $7E for WRAM access
+
+    ; CRITICAL: Save zero page variables on stack
+    ; This prevents corruption if NMI fires during game's block copy
+    ; Save $00-$0B (block copy uses $00-$06) and $26-$2B (Mult8 uses these)
+        lda.b 0x00
+        pha
+        lda.b 0x01
+        pha
+        lda.b 0x02
+        pha
+        lda.b 0x03
+        pha
+        lda.b 0x04
+        pha
+        lda.b 0x05
+        pha
+        lda.b 0x06
+        pha
+        lda.b 0x07
+        pha
+        lda.b 0x08
+        pha
+        lda.b 0x09
+        pha
+        lda.b 0x0A
+        pha
+        lda.b 0x0B
+        pha
+        ; Also save $26-$2B used by mult8_trampoline
+        lda.b 0x26
+        pha
+        lda.b 0x27
+        pha
+        lda.b 0x28
+        pha
+        lda.b 0x29
+        pha
+        lda.b 0x2A
+        pha
+        lda.b 0x2B
+        pha
+
+    ; Look up VRAM destination from slot table
+        lda.w rolling_slot_index
+        asl  ; ×2 for word lookup
+        tax
+        rep #0x20
+        lda.l _vram_slot_table, x  ; Get VRAM address ($7400, $7480, etc.)
+        sta.b 0x04  ; Save VRAM dest
+        sep #0x20
+
+    ; Calculate tilemap buffer source
+    ; source = tilemap_buffer_base + (slot × 128) - NO content_offset!
+        lda.w rolling_slot_index
+        sta.b 0x26
+        lda #TILEMAP_BYTES_PER_ROW
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
+        sta.b 0x00  ; Source address (low word)
+        sep #0x20
+
+    ; Queue using menu transfer system (entry 3 covers our slots)
+    ; The transfer will include this slot since it's at the right offset
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+
+        lda #0x01
+        sta.w 0x1825
+        sta.w 0x1824
+
+    ; CRITICAL: Restore zero page variables from stack (reverse order)
+    ; First restore $26-$2B (last pushed)
+        pla
+        sta.b 0x2B
+        pla
+        sta.b 0x2A
+        pla
+        sta.b 0x29
+        pla
+        sta.b 0x28
+        pla
+        sta.b 0x27
+        pla
+        sta.b 0x26
+        ; Then restore $00-$0B
+        pla
+        sta.b 0x0B
+        pla
+        sta.b 0x0A
+        pla
+        sta.b 0x09
+        pla
+        sta.b 0x08
+        pla
+        sta.b 0x07
+        pla
+        sta.b 0x06
+        pla
+        sta.b 0x05
+        pla
+        sta.b 0x04
+        pla
+        sta.b 0x03
+        pla
+        sta.b 0x02
+        pla
+        sta.b 0x01
+        pla
+        sta.b 0x00
+
+    ; Restore DBR
+        plb
+
+    ; Restore processor status (including interrupt flag)
+        plp
+        rts
+
+    ; ============================================================================
+    ; _copy_slot_to_tilemap
+    ; ============================================================================
+    ; Copies a single slot from text buffer to tilemap buffer
+    ; Input: rolling_slot_index = slot index (0-5)
+    ; Uses $00-$01, $26, $28, $2A, X, Y
+
+    _copy_slot_to_tilemap:
+        ; Save processor status and disable interrupts
+        ; Using PHP/PLP instead of SEI/CLI to handle nested calls correctly
+        php
+        sei
 
-    lda #0x01
-    sta.w 0x1825  ; 1 transfer only
-    sta.w 0x1824  ; Enable transfer
-
-    plb  ; Restore data bank
-    rtl
-
-; ============================================================================
-; _refresh_visible_items_internal
-; ============================================================================
-; Re-render all 5 visible items to our rolling buffer slots.
-; Uses current EF71 as the top item index.
-; Does NOT queue VRAM transfer (caller handles that).
-
-_refresh_visible_items_internal:
-    ; Render 5 visible items to their CORRECT circular buffer slots
-    ; The visible slots depend on rolling_buffer_pos due to circular rotation!
+    ; Save DBR - mult8_trampoline may change it
+        phb
+
+    ; Save and set D to 0 for direct page operations
+        rep #0x20
+        tdc
+        pha
+        lda.w #0x0000
+        tcd
+        sep #0x20
+
+    ; CRITICAL: Save zero page variables on stack
+    ; This prevents corruption if NMI fires during game's block copy
+    ; Save $00-$0B (block copy uses $00-$06) and $26-$2B (Mult8 uses these)
+        lda.b 0x00
+        pha
+        lda.b 0x01
+        pha
+        lda.b 0x02
+        pha
+        lda.b 0x03
+        pha
+        lda.b 0x04
+        pha
+        lda.b 0x05
+        pha
+        lda.b 0x06
+        pha
+        lda.b 0x07
+        pha
+        lda.b 0x08
+        pha
+        lda.b 0x09
+        pha
+        lda.b 0x0A
+        pha
+        lda.b 0x0B
+        pha
+        ; Also save $26-$2B used by mult8_trampoline
+        lda.b 0x26
+        pha
+        lda.b 0x27
+        pha
+        lda.b 0x28
+        pha
+        lda.b 0x29
+        pha
+        lda.b 0x2A
+        pha
+        lda.b 0x2B
+        pha
+
+    ; Calculate tilemap destination (NO content_offset - write to full slot)
+        lda.w rolling_slot_index
+        sta.b 0x26
+        lda #TILEMAP_BYTES_PER_ROW
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
+        sta.b 0x00
+        sep #0x20
+
+    ; Calculate text buffer source
+        lda.w rolling_slot_index
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        tax
+        sep #0x20
+
+    ; Copy row 1 (30 bytes) - content only, preserve borders
+        ldy.w #0x0000
+
+    _copy_slot_row1:
+        lda.w text_buffer_base, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #30
+        bne _copy_slot_row1
+
+    ; Copy row 2 (+$40 offset) - content only, preserve borders
+        ldy.w #0x0040
+
+    _copy_slot_row2:
+        lda.w text_buffer_base, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #0x005E
+        bne _copy_slot_row2
+
+    ; CRITICAL: Restore zero page variables $26-$2B from stack (reverse order - pushed last, pop first)
+        pla
+        sta.b 0x2B
+        pla
+        sta.b 0x2A
+        pla
+        sta.b 0x29
+        pla
+        sta.b 0x28
+        pla
+        sta.b 0x27
+        pla
+        sta.b 0x26
+
+    ; CRITICAL: Restore zero page variables $00-$0B from stack (reverse order)
+        pla
+        sta.b 0x0B
+        pla
+        sta.b 0x0A
+        pla
+        sta.b 0x09
+        pla
+        sta.b 0x08
+        pla
+        sta.b 0x07
+        pla
+        sta.b 0x06
+        pla
+        sta.b 0x05
+        pla
+        sta.b 0x04
+        pla
+        sta.b 0x03
+        pla
+        sta.b 0x02
+        pla
+        sta.b 0x01
+        pla
+        sta.b 0x00
+
+    ; Restore D register
+        rep #0x20
+        pla
+        tcd
+        sep #0x20
+
+    ; Restore DBR (pushed after php/sei)
+        plb
+
+    ; Restore processor status (including interrupt flag)
+        plp
+        rts
+
+    ; ============================================================================
+    ; _copy_all_slots_to_tilemap
+    ; ============================================================================
+    ; Copies all 6 slots from text buffer to tilemap buffer
+    ; Called from tfr_inventory_list_rolling AFTER game clears window buffers
+
+    _copy_all_slots_to_tilemap:
+        ; Save DBR - mult8_trampoline may change it
+        phb
+        lda #0x7E
+        pha
+        plb  ; Ensure DBR is $7E for WRAM access
+
+        lda #0
+        sta.b 0x06  ; Slot counter (0-5)
+
+    _copy_slots_loop:
+        ; Calculate tilemap destination: tilemap_buffer_base + (slot × 128)
+        ; NO content_offset - write to full 128-byte slot
+        lda.b 0x06
+        sta.b 0x26
+        lda #TILEMAP_BYTES_PER_ROW  ; 128
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        clc
+        adc #tilemap_buffer_base + tilemap_content_offset  ; NO +tilemap_content_offset!
+        sta.b 0x00  ; Tilemap dest pointer
+        sep #0x20
+
+    ; Calculate text buffer source: text_buffer_base + (slot × 60)
+        lda.b 0x06
+        sta.b 0x26
+        lda #TEXT_BYTES_PER_ITEM  ; 60
+        sta.b 0x28
+        jsr.l mult8_trampoline
+
+        rep #0x20
+        lda.b 0x2A
+        tax  ; X = text buffer offset
+        sep #0x20
+
+    ; Copy row 1 (30 bytes) - content only, preserve borders
+        ldy.w #0x0000
+
+    _copy_all_row1:
+        lda.w text_buffer_base, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #30
+        bne _copy_all_row1
+
+    ; Copy row 2 (+$40 offset in tilemap) - content only, preserve borders
+        ldy.w #0x0040
+
+    _copy_all_row2:
+        lda.w text_buffer_base, x
+        sta (0x00), y
+        inx
+        iny
+        cpy.w #0x005E
+        bne _copy_all_row2
+
+    ; Next slot
+        inc.b 0x06
+        lda.b 0x06
+        cmp #BUFFER_SLOTS  ; 6 slots
+        bne _copy_slots_loop
+
+    ; Restore DBR
+        plb
+        rts
+
+    ; ============================================================================
+    ; tfr_inventory_list_rolling
+    ; ============================================================================
+    ; Replacement for TfrInventoryList ($0298FA)
+    ; Transfers the visible portion of tilemap buffer to VRAM
     ;
+    ; Called via JSL from bank 02
 
-; After seam crossing, slots are rotated. For example if rolling_buffer_pos=3:
-;   - Slot 3 shows item at rolling_top_row + 0
-;   - Slot 4 shows item at rolling_top_row + 1
-;   - Slot 5 shows item at rolling_top_row + 2
-;   - Slot 0 shows item at rolling_top_row + 3 (wrapped)
-;   - Slot 1 shows item at rolling_top_row + 4
-
-    lda #0
-    sta.b 0x06  ; Visible row index (0-4)
-
-_refresh_int_loop:
-    ; Calculate item index = rolling_top_row + visible_row
-    lda.w rolling_top_row
-    clc
-    adc.b 0x06
-    sta.w rolling_edge_row  ; Item index for data lookup
-
-; Calculate actual slot = (rolling_buffer_pos + visible_row) % BUFFER_SLOTS
-    lda.w rolling_buffer_pos
-    clc
-    adc.b 0x06  ; pos + visible_row
-    cmp #BUFFER_SLOTS
-    bcc _refresh_slot_ok
-    sec
-    sbc #BUFFER_SLOTS  ; Wrap if >= 6
-
-_refresh_slot_ok:
-    sta.w rolling_slot_index  ; Actual circular buffer slot
-
-; Save visible row index
-    lda.b 0x06
-    pha
-
-; Render item to this slot
-    jsr.w _render_item_to_circular_slot
-
-; Restore visible row index
-    pla
-    sta.b 0x06
-
-; Next visible row
-    inc.b 0x06
-    lda.b 0x06
-    cmp #VISIBLE_ROWS  ; 5 visible items
-    bne _refresh_int_loop
-
-    rts
-
-; ============================================================================
-; _refresh_visible_items - Re-render all visible items after scroll
-; ============================================================================
-; Simpler approach: instead of circular buffer tricks, just redraw the 5
-; visible items to fixed slot positions (0-4) after each scroll.
-; EF65 is reset to 0 by the caller.
-;
-; Called via JSL from wrap_and_clear_trampoline
-
-_refresh_visible_items:
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Render 5 visible items to slots 0-4
-; Item index = EF71 + slot
-    lda #0
-    sta.b 0x06  ; Slot index (0-4)
-
-_refresh_loop:
-    ; Calculate item index = EF71 + slot
-    lda.w 0xEF71
-    clc
-    adc.b 0x06
-    sta.w rolling_edge_row  ; Item index for data lookup
-
-; Use slot index for buffer position
-    lda.b 0x06
-    sta.w rolling_slot_index  ; Slot index (0-4)
-
-; Save slot index
-    lda.b 0x06
-    pha
-
-; Render item to this slot
-    jsr.w _render_item_to_circular_slot
-
-; Restore slot index
-    pla
-    sta.b 0x06
-
-; Next slot
-    inc.b 0x06
-    lda.b 0x06
-    cmp #VISIBLE_ROWS  ; 5 visible items
-    bne _refresh_loop
-
-; Queue VRAM transfer
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-    lda #0x01
-    sta.w 0x1825
-    sta.w 0x1824
-
-    plb
-    rtl
-
-; ============================================================================
-; _post_render_up - Called after scroll UP animation completes
-; ============================================================================
-; Renders the next item (for potential next scroll UP) to the off-screen slot
-; Off-screen slot after scroll UP = (pos - 1) mod 6 (above top)
-;
-; Called via JSL from wrap_and_clear_trampoline
-
-_post_render_up:
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Calculate item index = EF71 + VISIBLE_ROWS (the next item down the list)
-    lda.w 0xEF71
-    clc
-    adc #VISIBLE_ROWS
-    cmp #TOTAL_ITEMS
-    bcs _post_up_done  ; Past end, nothing to render
-    sta.w rolling_edge_row
-
-; Off-screen slot = (pos - 1) mod 6
-    lda.w rolling_buffer_pos
-    dec
-    bpl _post_up_slot_ok
-    lda #BUFFER_SLOTS - 1
-
-_post_up_slot_ok:
-    sta.w rolling_slot_index
-
-    jsr.w _render_item_to_circular_slot
-
-; Queue VRAM transfer
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-    lda #0x01
-    sta.w 0x1825
-    sta.w 0x1824
-
-_post_up_done:
-    plb
-    rtl
-
-; ============================================================================
-; _post_render_down - Called after scroll DOWN animation completes
-; ============================================================================
-; Renders the previous item (for potential scroll UP back) to the off-screen slot
-; Off-screen slot after scroll DOWN = (pos + 5) mod 6 (below bottom)
-;
-; Called via JSL from wrap_and_clear_trampoline
-
-_post_render_down:
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Calculate item index = EF71 + VISIBLE_ROWS (the item just past visible bottom)
-; This prepares for scroll UP (which would show this item at bottom)
-    lda.w 0xEF71
-    clc
-    adc #VISIBLE_ROWS
-    cmp #TOTAL_ITEMS
-    bcs _post_down_done  ; Past end, nothing to render
-    sta.w rolling_edge_row
-
-; Off-screen slot = (pos + 5) mod 6
-    lda.w rolling_buffer_pos
-    clc
-    adc #VISIBLE_ROWS
-    cmp #BUFFER_SLOTS
-    bcc _post_down_slot_ok
-    sec
-    sbc #BUFFER_SLOTS
-
-_post_down_slot_ok:
-    sta.w rolling_slot_index
-
-    jsr.w _render_item_to_circular_slot
-
-; Queue VRAM transfer
-    lda #0x03
-    ldy.w #0x0002
-    jsr.l load_menu_tfr_data_trampoline
-    lda #0x01
-    sta.w 0x1825
-    sta.w 0x1824
-
-_post_down_done:
-    plb
-    rtl
-
-; ============================================================================
-; post_scroll_down_render
-; ============================================================================
-; Called via JSL from wrap_and_clear_trampoline after scroll animation completes.
-;
-; For scroll DOWN: Pre-render happened BEFORE animation (in scroll_list_down_hook)
-; so the newly visible slot already has correct content.
-;
-; We check if it was a scroll DOWN and optionally prepare the NEXT off-screen
-; slot for future scroll UP operations.
-
-post_scroll_down_render:
-"""Tail of `wrap_and_clear_trampoline` (JSL): post-animation hook for scroll-down."""
-    ; Check if this was a scroll DOWN animation (type 2)
-    ; $1820 still contains the animation type at this point
-    lda.w 0x1820
-    cmp #0x02
-    bne _psd_done
-
-; For scroll DOWN, the pre-render already happened in scroll_list_down_hook.
-; The slot that scrolled into view (at bottom) has correct content.
-;
-; Optionally, we could prepare the slot that's now off-screen (at top)
-; for a potential scroll UP. But since scroll_list_up_hook does pre-render,
-; this isn't strictly necessary.
-
-_psd_done:
-    rtl
-
-; ============================================================================
-; scroll_list_down_hook
-; ============================================================================
-; Called via JMP.L from $02A8B8
-; Starts scroll animation. Pre-rendering happens AFTER animation in post_scroll_down_render.
-; Returns via JMP.L to return_to_bank02 (which has RTS)
-;
-
-; Original code at $02A8B8:
-;   ldx $ef71, dex, stx $ef71, lda #$0c, sta $ef64, lda #$02, sta $1820, rts
-
-;MENU_FLAG_INVENTORY := 0x04         ; Bit 2 of $4A = inventory menu active
-
-scroll_list_down_hook:
-"""
-Replacement for the scroll-down stub at `$02A8B8`: kicks off the down-scroll animation, pre-renders the new bottom
-slot for the rolling buffer.
-"""
-; === GUARD: Only use rolling buffer for inventory menu ===
-; Check bit 2 of $4A (inventory flag). If not set, use original magic behavior.
-    lda.b 0x4A
-    and #0x04
-    bne _sd_is_inventory
-
-; --- ORIGINAL MAGIC MENU BEHAVIOR ---
-; Original code: ldx $ef71, dex, stx $ef71, lda #$0c, sta $ef64, lda #$02, sta $1820, rts
-    ldx.w 0xEF71
-    dex
-    stx.w 0xEF71
-    lda #0x0C
-    sta.w 0xEF64
-    lda #0x02
-    sta.w 0x1820
-    jmp.l return_to_bank02
-
-_sd_is_inventory:
-    ; NOTE: For inventory in single-column mode, we INCREMENT EF71 to show later items
+    tfr_inventory_list_rolling:
+    """
+    Replacement for original `TfrInventoryList` ($0298FA): re-render the visible inventory rows into our buffer and DMA
+    the slice to VRAM. Reached via JSL from bank 02.
+    """
+    ; Set data bank to $7E for WRAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; RE-RENDER all visible items only when something invalidated them
+    ; (open / item swap). Scroll-edge hooks paint the new hidden slot
+    ; before animation, so steady-state scroll requires zero re-render
+    ; here. The flag was the per-frame full rebuild that was clobbering
+    ; the pre-rendered hidden slot.
+        lda.w inventory_needs_full_refresh
+        beq _tfr_skip_refresh
+        jsr.w _refresh_visible_items_internal
+        stz.w inventory_needs_full_refresh
+
+    _tfr_skip_refresh:
+
+    ; Copy all 6 slots from text buffer to tilemap buffer
+    ; This runs AFTER the game's window clearing at $9AF4
+        jsr.w _copy_all_slots_to_tilemap
+
+    ; Queue VRAM transfer (entry 3 covers $7400-$77FF)
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+
+        lda #0x01
+        sta.w 0x1825  ; 1 transfer only
+        sta.w 0x1824  ; Enable transfer
+
+        plb  ; Restore data bank
+        rtl
+
+    ; ============================================================================
+    ; _refresh_visible_items_internal
+    ; ============================================================================
+    ; Re-render all 5 visible items to our rolling buffer slots.
+    ; Uses current EF71 as the top item index.
+    ; Does NOT queue VRAM transfer (caller handles that).
+
+    _refresh_visible_items_internal:
+        ; Render 5 visible items to their CORRECT circular buffer slots
+        ; The visible slots depend on rolling_buffer_pos due to circular rotation!
+        ;
+
+    ; After seam crossing, slots are rotated. For example if rolling_buffer_pos=3:
+    ;   - Slot 3 shows item at rolling_top_row + 0
+    ;   - Slot 4 shows item at rolling_top_row + 1
+    ;   - Slot 5 shows item at rolling_top_row + 2
+    ;   - Slot 0 shows item at rolling_top_row + 3 (wrapped)
+    ;   - Slot 1 shows item at rolling_top_row + 4
+
+        lda #0
+        sta.b 0x06  ; Visible row index (0-4)
+
+    _refresh_int_loop:
+        ; Calculate item index = rolling_top_row + visible_row
+        lda.w rolling_top_row
+        clc
+        adc.b 0x06
+        sta.w rolling_edge_row  ; Item index for data lookup
+
+    ; Calculate actual slot = (rolling_buffer_pos + visible_row) % BUFFER_SLOTS
+        lda.w rolling_buffer_pos
+        clc
+        adc.b 0x06  ; pos + visible_row
+        cmp #BUFFER_SLOTS
+        bcc _refresh_slot_ok
+        sec
+        sbc #BUFFER_SLOTS  ; Wrap if >= 6
+
+    _refresh_slot_ok:
+        sta.w rolling_slot_index  ; Actual circular buffer slot
+
+    ; Save visible row index
+        lda.b 0x06
+        pha
+
+    ; Render item to this slot
+        jsr.w _render_item_to_circular_slot
+
+    ; Restore visible row index
+        pla
+        sta.b 0x06
+
+    ; Next visible row
+        inc.b 0x06
+        lda.b 0x06
+        cmp #VISIBLE_ROWS  ; 5 visible items
+        bne _refresh_int_loop
+
+        rts
+
+    ; ============================================================================
+    ; _refresh_visible_items - Re-render all visible items after scroll
+    ; ============================================================================
+    ; Simpler approach: instead of circular buffer tricks, just redraw the 5
+    ; visible items to fixed slot positions (0-4) after each scroll.
+    ; EF65 is reset to 0 by the caller.
     ;
-    ; FF6-STYLE: Pre-render the new bottom item BEFORE starting the animation!
-    ; This ensures the slot has correct content when it scrolls into view.
-
-; Disable interrupts to prevent NMI from corrupting state
-    sei
-
-; Set data bank to $7E for RAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Check if we can scroll further: rolling_top_row + VISIBLE_ROWS must be < TOTAL_ITEMS
-; Use rolling_top_row, NOT EF71 which has different meaning in battle!
-    lda.w rolling_top_row
-    cmp #( TOTAL_ITEMS - VISIBLE_ROWS )
-    bcs _sd_abort  ; Already at max, can't scroll down
-
-; === FF6-STYLE PRE-RENDER ===
-; Before animation, render the item that will appear at the NEW bottom.
-; The "off-screen" slot below the visible area will scroll into view.
-; Off-screen slot = (rolling_buffer_pos + VISIBLE_ROWS) % BUFFER_SLOTS
-
-; Calculate item index = rolling_top_row + VISIBLE_ROWS (the item appearing at new bottom)
-; After increment, rolling_top_row will be current+1, so bottom item = (current+1) + 4 = current + 5
-    lda.w rolling_top_row
-    clc
-    adc #VISIBLE_ROWS  ; item = rolling_top_row + 5
-    cmp #TOTAL_ITEMS
-    bcs _sd_skip_prerender  ; Past end, nothing to render
-    sta.w rolling_edge_row  ; Save item index
-
-; Calculate the off-screen slot that will become visible
-; This is the slot 5 positions ahead of current top (wrapping)
-    lda.w rolling_buffer_pos
-    clc
-    adc #VISIBLE_ROWS  ; pos + 5
-    cmp #BUFFER_SLOTS
-    bcc _sd_slot_ok
-    sec
-    sbc #BUFFER_SLOTS  ; Wrap if >= 6
-
-_sd_slot_ok:
-    sta.w rolling_slot_index
-
-; Render item to the off-screen slot
-    jsr.w _render_item_to_circular_slot
-    ; Copy to tilemap buffer
-    jsr.w _copy_slot_to_tilemap
-    ; Queue VRAM transfer
-    jsr.w _transfer_circular_slot
-
-_sd_skip_prerender:
-    ; INCREMENT EF71 to show later items (animation loop is NOPed out)
-    lda.w 0xEF71
-    inc
-    sta.w 0xEF71
-
-; INCREMENT rolling_top_row to track which item is at top of visible area
-; This is critical for _refresh_visible_items_internal to work correctly!
-    lda.w rolling_top_row
-    inc
-    sta.w rolling_top_row
-
-; Set up scroll animation
-    lda #0x0C
-    sta.w 0xEF64
-    lda #0x02  ; Animation 2 (scroll down)
-    sta.w 0x1820
-
-; Advance circular buffer position (top slot moves up, becomes off-screen)
-    lda.w rolling_buffer_pos
-    inc
-    cmp #BUFFER_SLOTS
-    bcc _sd_pos_ok
-    lda #0
+    ; Called via JSL from wrap_and_clear_trampoline
+
+    _refresh_visible_items:
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Render 5 visible items to slots 0-4
+    ; Item index = EF71 + slot
+        lda #0
+        sta.b 0x06  ; Slot index (0-4)
+
+    _refresh_loop:
+        ; Calculate item index = EF71 + slot
+        lda.w 0xEF71
+        clc
+        adc.b 0x06
+        sta.w rolling_edge_row  ; Item index for data lookup
+
+    ; Use slot index for buffer position
+        lda.b 0x06
+        sta.w rolling_slot_index  ; Slot index (0-4)
+
+    ; Save slot index
+        lda.b 0x06
+        pha
+
+    ; Render item to this slot
+        jsr.w _render_item_to_circular_slot
+
+    ; Restore slot index
+        pla
+        sta.b 0x06
+
+    ; Next slot
+        inc.b 0x06
+        lda.b 0x06
+        cmp #VISIBLE_ROWS  ; 5 visible items
+        bne _refresh_loop
 
-_sd_pos_ok:
-    sta.w rolling_buffer_pos
+    ; Queue VRAM transfer
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+        lda #0x01
+        sta.w 0x1825
+        sta.w 0x1824
+
+        plb
+        rtl
+
+    ; ============================================================================
+    ; _post_render_up - Called after scroll UP animation completes
+    ; ============================================================================
+    ; Renders the next item (for potential next scroll UP) to the off-screen slot
+    ; Off-screen slot after scroll UP = (pos - 1) mod 6 (above top)
+    ;
+    ; Called via JSL from wrap_and_clear_trampoline
 
-_sd_abort:
-    ; Ensure cursor 1 stays visible during scroll animation
-    stz.w 0xEF69  ; Clear hide cursor 1 flag
-    stz.w 0xEF6E  ; Clear alternate hide cursor 1 flag
+    _post_render_up:
+        phb
+        lda #0x7E
+        pha
+        plb
 
-    plb  ; Restore data bank
-    cli  ; re-enable interrupts
-    jmp.l return_to_bank02
+    ; Calculate item index = EF71 + VISIBLE_ROWS (the next item down the list)
+        lda.w 0xEF71
+        clc
+        adc #VISIBLE_ROWS
+        cmp #TOTAL_ITEMS
+        bcs _post_up_done  ; Past end, nothing to render
+        sta.w rolling_edge_row
 
-; ============================================================================
-; scroll_list_up_hook
-; ============================================================================
-; Called via JMP.L from $02A8CA
-; Pre-renders new top row, then does original scroll setup
-; Returns via JMP.L to return_to_bank02 (which has RTS)
-;
+    ; Off-screen slot = (pos - 1) mod 6
+        lda.w rolling_buffer_pos
+        dec
+        bpl _post_up_slot_ok
+        lda #BUFFER_SLOTS - 1
 
-; Original code at $02A8CA:
-;   ldx $ef71, inx, stx $ef71, lda #$0c, sta $ef64, lda #$03, sta $1820, rts
+    _post_up_slot_ok:
+        sta.w rolling_slot_index
 
-scroll_list_up_hook:
-"""
-Replacement for the scroll-up stub at `$02A8CA`: pre-renders the new top row before kicking off the scroll-up
-animation.
-"""
-; === GUARD: Only use rolling buffer for inventory menu ===
-; Check bit 2 of $4A (inventory flag). If not set, use original magic behavior.
-    lda.b 0x4A
-    and #0x04
-    bne _su_is_inventory
-
-; --- ORIGINAL MAGIC MENU BEHAVIOR ---
-; Original code: ldx $ef71, inx, stx $ef71, lda #$0c, sta $ef64, lda #$03, sta $1820, rts
-    ldx.w 0xEF71
-    inx
-    stx.w 0xEF71
-    lda #0x0C
-    sta.w 0xEF64
-    lda #0x03
-    sta.w 0x1820
-    jmp.l return_to_bank02
+        jsr.w _render_item_to_circular_slot
 
-_su_is_inventory:
-    ; NOTE: For inventory in single-column mode, we DECREMENT rolling_top_row to show earlier items
+    ; Queue VRAM transfer
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+        lda #0x01
+        sta.w 0x1825
+        sta.w 0x1824
+
+    _post_up_done:
+        plb
+        rtl
+
+    ; ============================================================================
+    ; _post_render_down - Called after scroll DOWN animation completes
+    ; ============================================================================
+    ; Renders the previous item (for potential scroll UP back) to the off-screen slot
+    ; Off-screen slot after scroll DOWN = (pos + 5) mod 6 (below bottom)
     ;
+    ; Called via JSL from wrap_and_clear_trampoline
+
+    _post_render_down:
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Calculate item index = EF71 + VISIBLE_ROWS (the item just past visible bottom)
+    ; This prepares for scroll UP (which would show this item at bottom)
+        lda.w 0xEF71
+        clc
+        adc #VISIBLE_ROWS
+        cmp #TOTAL_ITEMS
+        bcs _post_down_done  ; Past end, nothing to render
+        sta.w rolling_edge_row
+
+    ; Off-screen slot = (pos + 5) mod 6
+        lda.w rolling_buffer_pos
+        clc
+        adc #VISIBLE_ROWS
+        cmp #BUFFER_SLOTS
+        bcc _post_down_slot_ok
+        sec
+        sbc #BUFFER_SLOTS
+
+    _post_down_slot_ok:
+        sta.w rolling_slot_index
+
+        jsr.w _render_item_to_circular_slot
 
-; For scroll UP, we DO need to pre-render BEFORE animation because:
-; - The slot that will scroll INTO view might have stale data from a previous scroll down
-; - We need to put the correct item there before it becomes visible
-
-; Disable interrupts to prevent NMI from corrupting rolling_slot_index
-    sei
-
-; Set data bank to $7E for RAM access
-    phb
-    lda #0x7E
-    pha
-    plb
-
-; Check if we can scroll: rolling_top_row must be > 0
-; (Use our custom variable, NOT EF71 which has different meaning in battle!)
-    lda.w rolling_top_row
-    beq _su_abort  ; Already at item 0, can't scroll up
-
-; First decrement buffer position (new top slot)
-    lda.w rolling_buffer_pos
-    dec
-    bpl _su_pos_ok
-    lda #BUFFER_SLOTS - 1  ; Wrap 0→5
-
-_su_pos_ok:
-    sta.w rolling_buffer_pos
-    sta.w rolling_slot_index  ; This slot will be the new top (currently off-screen)
-
-; Calculate previous item index = rolling_top_row - 1 (the item that will appear at top)
-; Use rolling_top_row, NOT EF71 which has different meaning in battle!
-    lda.w rolling_top_row
-    dec
-    sta.w rolling_edge_row
-
-; Pre-render to new top slot (currently off-screen, about to scroll in)
-    jsr.w _render_item_to_circular_slot
-    ; Copy to tilemap buffer
-    jsr.w _copy_slot_to_tilemap
-    ; Queue VRAM transfer for this slot
-    jsr.w _transfer_circular_slot
-
-; DECREMENT rolling_top_row to track which item is at top of visible area
-; This is the authoritative source for which items are visible!
-    lda.w rolling_top_row
-    dec
-    sta.w rolling_top_row
-
-; Set up scroll animation
-    lda #0x0C
-    sta.w 0xEF64
-    lda #0x03  ; Animation 3 (scroll up)
-    sta.w 0x1820
-
-_su_abort:
-    ; Ensure cursor 1 stays visible during scroll animation
-    stz.w 0xEF69  ; Clear hide cursor 1 flag
-    stz.w 0xEF6E  ; Clear alternate hide cursor 1 flag
-
-    plb  ; Restore data bank
-    cli  ; re-enable interrupts
-    jmp.l return_to_bank02
-
-; ============================================================================
-; update_list_scroll_hdma_wrapped
-; ============================================================================
-; Replacement for $02A7F1 - builds HDMA table with scroll value WRAPPING
-; This is the key to the circular buffer - scroll values wrap at 96 pixels
-;
-; Called via JMP.L from $02A7F1
-; Input: X = current scroll offset (from $EF65)
-;        Y = scanlines per row (12)
-; Output: HDMA table at $7F74 filled with wrapped scroll values
-
-HDMA_TABLE := 0x7E7F74  ; Full 24-bit address: bank $7E, offset $7F74 (V scroll entries)
-HDMA_TABLE_SIZE := 0x00F0  ; 240 bytes (same as HDMA_Y_SIZE)
-SCROLL_WRAP := 0x0060  ; 96 = 6 slots × 16 pixels (tilemap buffer size)
-SCROLL_WRAP_LIMIT := 0x01D3  ; 467 = 371 + 96 (wrap when >= this value)
-OUR_BASE_SCROLL := 0x0173  ; 371 - same as original game
-
-update_list_scroll_hdma_wrapped:
-"""
-Replacement for the HDMA-build stub at `$02A7F1`: rebuild the per-scanline V-scroll table for the inventory rolling
-buffer.
-"""
-; Check if we're in inventory mode (bit 2 of $4A)
-; If not, use original behavior for other windows
-    lda.b 0x4A
-    and #0x04
-    bne _use_circular_buffer
-
-; Original behavior for non-inventory windows
-; Input: X = scroll offset, Y = scanlines per row (12)
-    rep #0x20  ; 16-bit A
-    txa  ; A = scroll offset
-    ldx.w #0x0000  ; Table index
-
-_orig_loop:
-    sta.w 0x7F74, x  ; Store scroll value
-    dey
-    bne _orig_skip_add
-    clc
-    adc.w #0x0004  ; Add 4 every 12 scanlines
-    ldy.w #0x000C  ; Reset scanline counter
-
-_orig_skip_add:
-    inx
-    inx
-    inx
-    inx
-    cpx.w #0x00F0  ; 240 bytes
-    bne _orig_loop
-    .db 0x7B  ; TDC (clear A)
-    sep #0x20  ; 8-bit A
-    jmp.l return_to_bank02  ; Return via bank $02 trampoline
-
-_use_circular_buffer:
-    ; FF6-STYLE CIRCULAR BUFFER HDMA
+    ; Queue VRAM transfer
+        lda #0x03
+        ldy.w #0x0002
+        jsr.l load_menu_tfr_data_trampoline
+        lda #0x01
+        sta.w 0x1825
+        sta.w 0x1824
+
+    _post_down_done:
+        plb
+        rtl
+
+    ; ============================================================================
+    ; post_scroll_down_render
+    ; ============================================================================
+    ; Called via JSL from wrap_and_clear_trampoline after scroll animation completes.
+    ;
+    ; For scroll DOWN: Pre-render happened BEFORE animation (in scroll_list_down_hook)
+    ; so the newly visible slot already has correct content.
+    ;
+    ; We check if it was a scroll DOWN and optionally prepare the NEXT off-screen
+    ; slot for future scroll UP operations.
+
+    post_scroll_down_render:
+    """Tail of `wrap_and_clear_trampoline` (JSL): post-animation hook for scroll-down."""
+        ; Check if this was a scroll DOWN animation (type 2)
+        ; $1820 still contains the animation type at this point
+        lda.w 0x1820
+        cmp #0x02
+        bne _psd_done
+
+    ; For scroll DOWN, the pre-render already happened in scroll_list_down_hook.
+    ; The slot that scrolled into view (at bottom) has correct content.
+    ;
+    ; Optionally, we could prepare the slot that's now off-screen (at top)
+    ; for a potential scroll UP. But since scroll_list_up_hook does pre-render,
+    ; this isn't strictly necessary.
+
+    _psd_done:
+        rtl
+
+    ; ============================================================================
+    ; scroll_list_down_hook
+    ; ============================================================================
+    ; Called via JMP.L from $02A8B8
+    ; Starts scroll animation. Pre-rendering happens AFTER animation in post_scroll_down_render.
+    ; Returns via JMP.L to return_to_bank02 (which has RTS)
     ;
 
-; Key insight from FF6's LoadItemBG1VScrollHDMATbl:
-;   - At screen scanline S with scroll V, displayed VRAM row = V + S
-
-;   - To show VRAM slot N at screen row R (scanline R*12):
-;     scroll + R*12 = N*16, therefore scroll = N*16 - R*12
-;
-
-; Algorithm for each visible row (0-4):
-;   vram_slot = (rolling_buffer_pos + row) % 6
-;   scroll = BASE + (vram_slot * 16) - (row * 12)
-;
-; This ensures proper alignment AND seamless wraparound.
-;
-    sei  ; Disable interrupts during HDMA table update
-
-; CRITICAL: Save direct page variables that the caller depends on
-    lda.b 0x00
-    pha
-    lda.b 0x01
-    pha
-    lda.b 0x02
-    pha
-    lda.b 0x03
-    pha
-    lda.b 0x04
-    pha
-    lda.b 0x05
-    pha
-
-    rep #0x30  ; 16-bit A, X, Y
-
-    ldx.w #0x0000  ; HDMA table index
-    stz.b 0x02  ; Row counter (clears $02 and $03 in 16-bit mode)
-
-_row_loop:
-    ; Calculate vram_slot = (rolling_buffer_pos + row) % 6
-    lda.w rolling_buffer_pos
-    and.w #0x00FF
-    clc
-    adc.b 0x02  ; + row number
-
-_mod6:
-    cmp.w #BUFFER_SLOTS  ; >= 6?
-    bcc _mod6_done
-    sec
-    sbc.w #BUFFER_SLOTS
-    bra _mod6
-
-_mod6_done:
-    ; A = vram_slot (0-5)
-
-; Calculate vram_slot * 16
-    asl
-    asl
-    asl
-    asl  ; A = vram_slot * 16
-    sta.b 0x00  ; Save vram_offset
-
-; Calculate row * 12 = row * 8 + row * 4
-; Use Y register for temp to avoid clobbering direct page
-    lda.b 0x02  ; row (only low byte matters, high is 0)
-    and.w #0x00FF  ; Ensure only low byte
-    asl
-    asl
-    asl  ; row * 8
-    tay  ; Y = row * 8
-    lda.b 0x02
-    and.w #0x00FF
-    asl
-    asl  ; row * 4
-    sta.b 0x04  ; Use $04 for temp (not overlapping)
-    tya  ; A = row * 8
-    clc
-    adc.b 0x04  ; row * 8 + row * 4 = row * 12
-    ; A = scanline_offset (row * 12)
-
-; scroll = BASE + vram_offset - scanline_offset
-; eor.w   #0xFFFF                 ; Negate: -scanline_offset
-; a816 does not support yet the w for eor.
-    .db 0x49
-    .dw 0xffff
-
-    inc
-    clc
-    adc.b 0x00  ; + vram_offset
-    clc
-    adc.w #OUR_BASE_SCROLL  ; + BASE
-    sta.b 0x00  ; $00 = scroll value for this row
-
-; Store same scroll value for all 12 scanlines of this row
-    ldy.w #0x000C  ; 12 scanlines
-
-_scanline_loop:
-    lda.b 0x00
-    sta.l HDMA_TABLE, x
-    inx
-    inx
-    inx
-    inx  ; X += 4 (next HDMA entry)
-    dey
-    bne _scanline_loop
-
-; Next row
-    inc.b 0x02
-    lda.b 0x02
-    cmp.w #VISIBLE_ROWS  ; 5 rows total
-    bne _row_loop
+    ; Original code at $02A8B8:
+    ;   ldx $ef71, dex, stx $ef71, lda #$0c, sta $ef64, lda #$02, sta $1820, rts
+
+    ;MENU_FLAG_INVENTORY := 0x04         ; Bit 2 of $4A = inventory menu active
+
+    scroll_list_down_hook:
+    """
+    Replacement for the scroll-down stub at `$02A8B8`: kicks off the down-scroll animation, pre-renders the new bottom
+    slot for the rolling buffer.
+    """
+    ; === GUARD: Only use rolling buffer for inventory menu ===
+    ; Check bit 2 of $4A (inventory flag). If not set, use original magic behavior.
+        lda.b 0x4A
+        and #0x04
+        bne _sd_is_inventory
+
+    ; --- ORIGINAL MAGIC MENU BEHAVIOR ---
+    ; Original code: ldx $ef71, dex, stx $ef71, lda #$0c, sta $ef64, lda #$02, sta $1820, rts
+        ldx.w 0xEF71
+        dex
+        stx.w 0xEF71
+        lda #0x0C
+        sta.w 0xEF64
+        lda #0x02
+        sta.w 0x1820
+        jmp.l return_to_bank02
+
+    _sd_is_inventory:
+        ; NOTE: For inventory in single-column mode, we INCREMENT EF71 to show later items
+        ;
+        ; FF6-STYLE: Pre-render the new bottom item BEFORE starting the animation!
+        ; This ensures the slot has correct content when it scrolls into view.
+
+    ; Disable interrupts to prevent NMI from corrupting state
+        sei
+
+    ; Set data bank to $7E for RAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Check if we can scroll further: rolling_top_row + VISIBLE_ROWS must be < TOTAL_ITEMS
+    ; Use rolling_top_row, NOT EF71 which has different meaning in battle!
+        lda.w rolling_top_row
+        cmp #( TOTAL_ITEMS - VISIBLE_ROWS )
+        bcs _sd_abort  ; Already at max, can't scroll down
+
+    ; === FF6-STYLE PRE-RENDER ===
+    ; Before animation, render the item that will appear at the NEW bottom.
+    ; The "off-screen" slot below the visible area will scroll into view.
+    ; Off-screen slot = (rolling_buffer_pos + VISIBLE_ROWS) % BUFFER_SLOTS
+
+    ; Calculate item index = rolling_top_row + VISIBLE_ROWS (the item appearing at new bottom)
+    ; After increment, rolling_top_row will be current+1, so bottom item = (current+1) + 4 = current + 5
+        lda.w rolling_top_row
+        clc
+        adc #VISIBLE_ROWS  ; item = rolling_top_row + 5
+        cmp #TOTAL_ITEMS
+        bcs _sd_skip_prerender  ; Past end, nothing to render
+        sta.w rolling_edge_row  ; Save item index
+
+    ; Calculate the off-screen slot that will become visible
+    ; This is the slot 5 positions ahead of current top (wrapping)
+        lda.w rolling_buffer_pos
+        clc
+        adc #VISIBLE_ROWS  ; pos + 5
+        cmp #BUFFER_SLOTS
+        bcc _sd_slot_ok
+        sec
+        sbc #BUFFER_SLOTS  ; Wrap if >= 6
+
+    _sd_slot_ok:
+        sta.w rolling_slot_index
+
+    ; Render item to the off-screen slot
+        jsr.w _render_item_to_circular_slot
+        ; Copy to tilemap buffer
+        jsr.w _copy_slot_to_tilemap
+        ; Queue VRAM transfer
+        jsr.w _transfer_circular_slot
+
+    _sd_skip_prerender:
+        ; INCREMENT EF71 to show later items (animation loop is NOPed out)
+        lda.w 0xEF71
+        inc
+        sta.w 0xEF71
+
+    ; INCREMENT rolling_top_row to track which item is at top of visible area
+    ; This is critical for _refresh_visible_items_internal to work correctly!
+        lda.w rolling_top_row
+        inc
+        sta.w rolling_top_row
+
+    ; Set up scroll animation
+        lda #0x0C
+        sta.w 0xEF64
+        lda #0x02  ; Animation 2 (scroll down)
+        sta.w 0x1820
+
+    ; Advance circular buffer position (top slot moves up, becomes off-screen)
+        lda.w rolling_buffer_pos
+        inc
+        cmp #BUFFER_SLOTS
+        bcc _sd_pos_ok
+        lda #0
+
+    _sd_pos_ok:
+        sta.w rolling_buffer_pos
+
+    _sd_abort:
+        ; Ensure cursor 1 stays visible during scroll animation
+        stz.w 0xEF69  ; Clear hide cursor 1 flag
+        stz.w 0xEF6E  ; Clear alternate hide cursor 1 flag
+
+        plb  ; Restore data bank
+        cli  ; re-enable interrupts
+        jmp.l return_to_bank02
+
+    ; ============================================================================
+    ; scroll_list_up_hook
+    ; ============================================================================
+    ; Called via JMP.L from $02A8CA
+    ; Pre-renders new top row, then does original scroll setup
+    ; Returns via JMP.L to return_to_bank02 (which has RTS)
+    ;
 
-; CRITICAL: Restore direct page variables before returning
-    sep #0x20  ; 8-bit A for PLA
-    pla
-    sta.b 0x05
-    pla
-    sta.b 0x04
-    pla
-    sta.b 0x03
-    pla
-    sta.b 0x02
-    pla
-    sta.b 0x01
-    pla
-    sta.b 0x00
+    ; Original code at $02A8CA:
+    ;   ldx $ef71, inx, stx $ef71, lda #$0c, sta $ef64, lda #$03, sta $1820, rts
+
+    scroll_list_up_hook:
+    """
+    Replacement for the scroll-up stub at `$02A8CA`: pre-renders the new top row before kicking off the scroll-up
+    animation.
+    """
+    ; === GUARD: Only use rolling buffer for inventory menu ===
+    ; Check bit 2 of $4A (inventory flag). If not set, use original magic behavior.
+        lda.b 0x4A
+        and #0x04
+        bne _su_is_inventory
+
+    ; --- ORIGINAL MAGIC MENU BEHAVIOR ---
+    ; Original code: ldx $ef71, inx, stx $ef71, lda #$0c, sta $ef64, lda #$03, sta $1820, rts
+        ldx.w 0xEF71
+        inx
+        stx.w 0xEF71
+        lda #0x0C
+        sta.w 0xEF64
+        lda #0x03
+        sta.w 0x1820
+        jmp.l return_to_bank02
+
+    _su_is_inventory:
+        ; NOTE: For inventory in single-column mode, we DECREMENT rolling_top_row to show earlier items
+        ;
+
+    ; For scroll UP, we DO need to pre-render BEFORE animation because:
+    ; - The slot that will scroll INTO view might have stale data from a previous scroll down
+    ; - We need to put the correct item there before it becomes visible
 
-; Check if scroll animation is active before forcing cursor position
-; Only force during animation ($1820 = 2 or 3), otherwise let normal code handle it
-    lda.l 0x7E1820  ; Animation type
-    beq _cursor_skip_force  ; If 0, no animation - skip forcing
-
-; Animation is active - ensure cursor stays visible
-; Must use long addressing since data bank may be $02 (ROM), not $7E (WRAM)
-    pha  ; Save animation type
-    lda #0x00
-    sta.l 0x7EEF69  ; Clear hide cursor 1 flag
-    sta.l 0x7EEF6E  ; Clear alternate hide cursor 1 flag
-
-; X position is always $0C for single-column mode
-    lda #0x0C
-    sta.l 0x7EEF6B  ; Set cursor 1 X position
-
-; Set Y based on animation direction
-; Scroll down ($1820=2): cursor at bottom row, Y = $CC
-; Scroll up ($1820=3): cursor at top row, Y = $9C
-    pla  ; Restore animation type
-    cmp #0x02  ; Scroll down?
-    bne _cursor_scroll_up
-    lda #0xCC  ; Bottom row Y position
-    bra _cursor_set_y
-
-_cursor_scroll_up:
-    lda #0x9C  ; Top row Y position
-
-_cursor_set_y:
-    sta.l 0x7EEF6D  ; Set cursor 1 Y position
-
-_cursor_skip_force:
-
-    cli
-    jmp.l return_to_bank02
-
-; HDMA table addresses - Y scroll portion only!
-; Original ResetListScrollHDMA writes to $81F4 (Y scroll in swap table)
-; Animation swaps this with $7F74 (Y scroll in active table)
-; X scroll values at $81D2/$7F52 are left alone.
-HDMA_SWAP_Y := 0x7E81F4  ; Y scroll in swap table
-HDMA_ACTIVE_Y := 0x7E7F74  ; Y scroll in active table
-HDMA_Y_SIZE := 0x00F0  ; 240 bytes (same as original)
-
-; ============================================================================
-; reset_list_scroll_hdma_rolling
-; ============================================================================
-; Replacement for $02AAB8 - fills Y scroll in BOTH HDMA tables
-; X scroll values stay as original game initialized them.
-;
-; Only difference from original: uses 132-based values instead of 371-based.
-
-reset_list_scroll_hdma_rolling:
-"""Replacement for `$02AAB8`: fill the Y-scroll bytes of both HDMA tables when the inventory window opens."""
+    ; Disable interrupts to prevent NMI from corrupting rolling_slot_index
+        sei
+
+    ; Set data bank to $7E for RAM access
+        phb
+        lda #0x7E
+        pha
+        plb
+
+    ; Check if we can scroll: rolling_top_row must be > 0
+    ; (Use our custom variable, NOT EF71 which has different meaning in battle!)
+        lda.w rolling_top_row
+        beq _su_abort  ; Already at item 0, can't scroll up
+
+    ; First decrement buffer position (new top slot)
+        lda.w rolling_buffer_pos
+        dec
+        bpl _su_pos_ok
+        lda #BUFFER_SLOTS - 1  ; Wrap 0→5
+
+    _su_pos_ok:
+        sta.w rolling_buffer_pos
+        sta.w rolling_slot_index  ; This slot will be the new top (currently off-screen)
+
+    ; Calculate previous item index = rolling_top_row - 1 (the item that will appear at top)
+    ; Use rolling_top_row, NOT EF71 which has different meaning in battle!
+        lda.w rolling_top_row
+        dec
+        sta.w rolling_edge_row
+
+    ; Pre-render to new top slot (currently off-screen, about to scroll in)
+        jsr.w _render_item_to_circular_slot
+        ; Copy to tilemap buffer
+        jsr.w _copy_slot_to_tilemap
+        ; Queue VRAM transfer for this slot
+        jsr.w _transfer_circular_slot
+
+    ; DECREMENT rolling_top_row to track which item is at top of visible area
+    ; This is the authoritative source for which items are visible!
+        lda.w rolling_top_row
+        dec
+        sta.w rolling_top_row
+
+    ; Set up scroll animation
+        lda #0x0C
+        sta.w 0xEF64
+        lda #0x03  ; Animation 3 (scroll up)
+        sta.w 0x1820
+        bra _su_exit
+
+    _su_abort:
+        ; Pre-increment EF85, EF86, $63 so the caller's unconditional
+        ; `dec EF86 / dec EF85 / dec $63` at $02:B4FA-B500 lands back on
+        ; the original values when our scroll-up aborts at the top of
+        ; the list. Without this, $63 (cursor item index) underflows
+        ; from 0 to 0xFF and the row 0 cursor item points at junk,
+        ; reproducing the "one extra UP past item 0" visual.
+        inc.w 0xEF85
+        inc.w 0xEF86
+        inc.b 0x63
+
+    _su_exit:
+        ; Ensure cursor 1 stays visible during scroll animation
+        stz.w 0xEF69  ; Clear hide cursor 1 flag
+        stz.w 0xEF6E  ; Clear alternate hide cursor 1 flag
+
+        plb  ; Restore data bank
+        cli  ; re-enable interrupts
+        jmp.l return_to_bank02
+
+    ; ============================================================================
+    ; update_list_scroll_hdma_wrapped
+    ; ============================================================================
+    ; Replacement for $02A7F1 - builds HDMA table with scroll value WRAPPING
+    ; This is the key to the circular buffer - scroll values wrap at 96 pixels
+    ;
+    ; Called via JMP.L from $02A7F1
+    ; Input: X = current scroll offset (from $EF65)
+    ;        Y = scanlines per row (12)
+    ; Output: HDMA table at $7F74 filled with wrapped scroll values
+
+    HDMA_TABLE := 0x7E7F74  ; Full 24-bit address: bank $7E, offset $7F74 (V scroll entries)
+    HDMA_TABLE_SIZE := 0x00F0  ; 240 bytes (same as HDMA_Y_SIZE)
+    SCROLL_WRAP := 0x0060  ; 96 = 6 slots × 16 pixels (tilemap buffer size)
+    SCROLL_WRAP_LIMIT := 0x01D3  ; 467 = 371 + 96 (wrap when >= this value)
+    OUR_BASE_SCROLL := 0x0173  ; 371 - same as original game
+
+    update_list_scroll_hdma_wrapped:
+    """
+    Replacement for the HDMA-build stub at `$02A7F1`: rebuild the per-scanline V-scroll table for the inventory rolling
+    buffer.
+    """
     ; Check if we're in inventory mode (bit 2 of $4A)
-    lda.b 0x4A
-    and #0x04
-    bne _reset_use_circular
-
-; Original behavior for non-inventory windows
-    ldx.w #0x0173  ; 371 - base scroll
-    stx.w 0xEF65
-    ldy.w #0x000C  ; 12 scanlines
-    sty.w 0xEF67
-    rep #0x20  ; 16-bit A
-    txa  ; A = 371
-    ldx.w #0x0000
-
-_reset_orig_loop:
-    sta.w 0x81F4, x  ; Swap table only (like original)
-    dey
-    bne _reset_orig_skip
-    clc
-    adc.w #0x0004  ; Add 4 every 12 scanlines
-    ldy.w #0x000C
-
-_reset_orig_skip:
-    inx
-    inx
-    inx
-    inx
-    cpx.w #0x00F0
-    bne _reset_orig_loop
-    .db 0x7B  ; TDC
-    sep #0x20
-    rtl
-
-_reset_use_circular:
-    ; Circular buffer setup for inventory
-    ; Save $00-$03 to prevent corrupting caller's state
-    lda.b 0x00
-    pha
-    lda.b 0x01
-    pha
-    lda.b 0x02
-    pha
-    lda.b 0x03
-    pha
-
-    ldx.w #0x0173  ; 371 - game's expected base scroll
-    stx.w 0xEF65
-    ldy.w #0x000C  ; 12 scanlines per item row
-    sty.w 0xEF67
-
-; Re-initialize circular buffer contents when inventory (re)opens
-; This ensures VRAM has correct items even after closing/reopening
-    jsr.l init_inventory_text_buf_rolling
-
-; rolling_buffer_pos is now set by init_inventory_text_buf_rolling
-
-; Fill Y scroll in both tables using circular buffer formula
-    stz.b 0x02  ; Row counter low byte (8-bit mode)
-    stz.b 0x03  ; Row counter high byte
-    rep #0x30  ; 16-bit A and X/Y
-    ldx.w #0x0000  ; HDMA table index
-
-_reset_row_loop:
-    ; At init: vram_slot = row, so scroll = BASE + row * 4
-    lda.b 0x02  ; row
-    and.w #0x00FF
-    asl
-    asl  ; row * 4
-    clc
-    adc.w #OUR_BASE_SCROLL  ; + BASE
-    sta.b 0x00  ; Save scroll value
-
-; Store same value for all 12 scanlines of this row
-    ldy.w #0x000C  ; 12 scanlines
-
-_reset_scanline_loop:
-    lda.b 0x00
-    sta.l HDMA_SWAP_Y, x  ; $81F4 + X
-    sta.l HDMA_ACTIVE_Y, x  ; $7F74 + X
-    inx
-    inx
-    inx
-    inx  ; X += 4
-    dey
-    bne _reset_scanline_loop
-
-; Next row
-    inc.b 0x02
-    lda.b 0x02
-    cmp.w #VISIBLE_ROWS  ; 5 rows
-    bne _reset_row_loop
-
-    sep #0x20  ; 8-bit A (game expects this)
-    ; Restore $00-$03 before returning
-    pla
-    sta.b 0x03
-    pla
-    sta.b 0x02
-    pla
-    sta.b 0x01
-    pla
-    sta.b 0x00
-    rtl
-
-; ============================================================================
-; check_cursor2_visibility_rolling
-; ============================================================================
-; Checks if cursor 2 (the first selected item in swap mode) should be visible.
-; Called after scroll animation completes.
-;
-
-; Logic:
-;   - If swap mode NOT active ($EF94 == 0), return immediately
-;   - Get first selected item index from $EF95 (mask out bit 7)
+    ; If not, use original behavior for other windows
+        lda.b 0x4A
+        and #0x04
+        bne _use_circular_buffer
+
+    ; Original behavior for non-inventory windows
+    ; Input: X = scroll offset, Y = scanlines per row (12)
+        rep #0x20  ; 16-bit A
+        txa  ; A = scroll offset
+        ldx.w #0x0000  ; Table index
+
+    _orig_loop:
+        sta.w 0x7F74, x  ; Store scroll value
+        dey
+        bne _orig_skip_add
+        clc
+        adc.w #0x0004  ; Add 4 every 12 scanlines
+        ldy.w #0x000C  ; Reset scanline counter
+
+    _orig_skip_add:
+        inx
+        inx
+        inx
+        inx
+        cpx.w #0x00F0  ; 240 bytes
+        bne _orig_loop
+        .db 0x7B  ; TDC (clear A)
+        sep #0x20  ; 8-bit A
+        jmp.l return_to_bank02  ; Return via bank $02 trampoline
+
+    _use_circular_buffer:
+        ; FF6-STYLE CIRCULAR BUFFER HDMA
+        ;
+
+    ; Key insight from FF6's LoadItemBG1VScrollHDMATbl:
+    ;   - At screen scanline S with scroll V, displayed VRAM row = V + S
+
+    ;   - To show VRAM slot N at screen row R (scanline R*12):
+    ;     scroll + R*12 = N*16, therefore scroll = N*16 - R*12
+    ;
 
-;   - If rolling_top_row <= selected < rolling_top_row + VISIBLE_ROWS:
-;       Show cursor 2 ($EF6A = 0)
+    ; Algorithm for each visible row (0-4):
+    ;   vram_slot = (rolling_buffer_pos + row) % 6
+    ;   scroll = BASE + (vram_slot * 16) - (row * 12)
+    ;
+    ; This ensures proper alignment AND seamless wraparound.
+    ;
+        sei  ; Disable interrupts during HDMA table update
+
+    ; CRITICAL: Save direct page variables that the caller depends on
+        lda.b 0x00
+        pha
+        lda.b 0x01
+        pha
+        lda.b 0x02
+        pha
+        lda.b 0x03
+        pha
+        lda.b 0x04
+        pha
+        lda.b 0x05
+        pha
+
+        rep #0x30  ; 16-bit A, X, Y
+
+        ldx.w #0x0000  ; HDMA table index
+        stz.b 0x02  ; Row counter (clears $02 and $03 in 16-bit mode)
+
+    _row_loop:
+        ; Calculate vram_slot = (rolling_buffer_pos + row) % 6
+        lda.w rolling_buffer_pos
+        and.w #0x00FF
+        clc
+        adc.b 0x02  ; + row number
+
+    _mod6:
+        cmp.w #BUFFER_SLOTS  ; >= 6?
+        bcc _mod6_done
+        sec
+        sbc.w #BUFFER_SLOTS
+        bra _mod6
+
+    _mod6_done:
+        ; A = vram_slot (0-5)
+
+    ; Calculate vram_slot * 16
+        asl
+        asl
+        asl
+        asl  ; A = vram_slot * 16
+        sta.b 0x00  ; Save vram_offset
+
+    ; Calculate row * 12 = row * 8 + row * 4
+    ; Use Y register for temp to avoid clobbering direct page
+        lda.b 0x02  ; row (only low byte matters, high is 0)
+        and.w #0x00FF  ; Ensure only low byte
+        asl
+        asl
+        asl  ; row * 8
+        tay  ; Y = row * 8
+        lda.b 0x02
+        and.w #0x00FF
+        asl
+        asl  ; row * 4
+        sta.b 0x04  ; Use $04 for temp (not overlapping)
+        tya  ; A = row * 8
+        clc
+        adc.b 0x04  ; row * 8 + row * 4 = row * 12
+        ; A = scanline_offset (row * 12)
+
+    ; scroll = BASE + vram_offset - scanline_offset
+    ; eor.w   #0xFFFF                 ; Negate: -scanline_offset
+    ; a816 does not support yet the w for eor.
+        .db 0x49
+        .dw 0xffff
+
+        inc
+        clc
+        adc.b 0x00  ; + vram_offset
+        clc
+        adc.w #OUR_BASE_SCROLL  ; + BASE
+        sta.b 0x00  ; $00 = scroll value for this row
+
+    ; Store same scroll value for all 12 scanlines of this row
+        ldy.w #0x000C  ; 12 scanlines
+
+    _scanline_loop:
+        lda.b 0x00
+        sta.l HDMA_TABLE, x
+        inx
+        inx
+        inx
+        inx  ; X += 4 (next HDMA entry)
+        dey
+        bne _scanline_loop
+
+    ; Next row
+        inc.b 0x02
+        lda.b 0x02
+        cmp.w #VISIBLE_ROWS  ; 5 rows total
+        bne _row_loop
+
+    ; CRITICAL: Restore direct page variables before returning
+        sep #0x20  ; 8-bit A for PLA
+        pla
+        sta.b 0x05
+        pla
+        sta.b 0x04
+        pla
+        sta.b 0x03
+        pla
+        sta.b 0x02
+        pla
+        sta.b 0x01
+        pla
+        sta.b 0x00
+
+    ; Check if scroll animation is active before forcing cursor position
+    ; Only force during animation ($1820 = 2 or 3), otherwise let normal code handle it
+        lda.l 0x7E1820  ; Animation type
+        beq _cursor_skip_force  ; If 0, no animation - skip forcing
+
+    ; Animation is active - ensure cursor stays visible
+    ; Must use long addressing since data bank may be $02 (ROM), not $7E (WRAM)
+        pha  ; Save animation type
+        lda #0x00
+        sta.l 0x7EEF69  ; Clear hide cursor 1 flag
+        sta.l 0x7EEF6E  ; Clear alternate hide cursor 1 flag
+
+    ; X position is always $0C for single-column mode
+        lda #0x0C
+        sta.l 0x7EEF6B  ; Set cursor 1 X position
+
+    ; Set Y based on animation direction
+    ; Scroll down ($1820=2): cursor at bottom row, Y = $CC
+    ; Scroll up ($1820=3): cursor at top row, Y = $9C
+        pla  ; Restore animation type
+        cmp #0x02  ; Scroll down?
+        bne _cursor_scroll_up
+        lda #0xCC  ; Bottom row Y position
+        bra _cursor_set_y
+
+    _cursor_scroll_up:
+        lda #0x9C  ; Top row Y position
+
+    _cursor_set_y:
+        sta.l 0x7EEF6D  ; Set cursor 1 Y position
+
+    _cursor_skip_force:
+
+        cli
+        jmp.l return_to_bank02
+
+    ; HDMA table addresses - Y scroll portion only!
+    ; Original ResetListScrollHDMA writes to $81F4 (Y scroll in swap table)
+    ; Animation swaps this with $7F74 (Y scroll in active table)
+    ; X scroll values at $81D2/$7F52 are left alone.
+    HDMA_SWAP_Y := 0x7E81F4  ; Y scroll in swap table
+    HDMA_ACTIVE_Y := 0x7E7F74  ; Y scroll in active table
+    HDMA_Y_SIZE := 0x00F0  ; 240 bytes (same as original)
+
+    ; ============================================================================
+    ; reset_list_scroll_hdma_rolling
+    ; ============================================================================
+    ; Replacement for $02AAB8 - fills Y scroll in BOTH HDMA tables
+    ; X scroll values stay as original game initialized them.
+    ;
+    ; Only difference from original: uses 132-based values instead of 371-based.
+
+    reset_list_scroll_hdma_rolling:
+    """Replacement for `$02AAB8`: fill the Y-scroll bytes of both HDMA tables when the inventory window opens."""
+        ; Check if we're in inventory mode (bit 2 of $4A)
+        lda.b 0x4A
+        and #0x04
+        bne _reset_use_circular
+
+    ; Original behavior for non-inventory windows
+        ldx.w #0x0173  ; 371 - base scroll
+        stx.w 0xEF65
+        ldy.w #0x000C  ; 12 scanlines
+        sty.w 0xEF67
+        rep #0x20  ; 16-bit A
+        txa  ; A = 371
+        ldx.w #0x0000
+
+    _reset_orig_loop:
+        sta.w 0x81F4, x  ; Swap table only (like original)
+        dey
+        bne _reset_orig_skip
+        clc
+        adc.w #0x0004  ; Add 4 every 12 scanlines
+        ldy.w #0x000C
+
+    _reset_orig_skip:
+        inx
+        inx
+        inx
+        inx
+        cpx.w #0x00F0
+        bne _reset_orig_loop
+        .db 0x7B  ; TDC
+        sep #0x20
+        rtl
+
+    _reset_use_circular:
+        ; Circular buffer setup for inventory
+        ; Save $00-$03 to prevent corrupting caller's state
+        lda.b 0x00
+        pha
+        lda.b 0x01
+        pha
+        lda.b 0x02
+        pha
+        lda.b 0x03
+        pha
+
+        ldx.w #0x0173  ; 371 - game's expected base scroll
+        stx.w 0xEF65
+        ldy.w #0x000C  ; 12 scanlines per item row
+        sty.w 0xEF67
+
+    ; Re-initialize circular buffer contents when inventory (re)opens
+    ; This ensures VRAM has correct items even after closing/reopening
+        jsr.l init_inventory_text_buf_rolling
+
+    ; rolling_buffer_pos is now set by init_inventory_text_buf_rolling
+
+    ; Fill Y scroll in both tables using circular buffer formula
+        stz.b 0x02  ; Row counter low byte (8-bit mode)
+        stz.b 0x03  ; Row counter high byte
+        rep #0x30  ; 16-bit A and X/Y
+        ldx.w #0x0000  ; HDMA table index
+
+    _reset_row_loop:
+        ; At init: vram_slot = row, so scroll = BASE + row * 4
+        lda.b 0x02  ; row
+        and.w #0x00FF
+        asl
+        asl  ; row * 4
+        clc
+        adc.w #OUR_BASE_SCROLL  ; + BASE
+        sta.b 0x00  ; Save scroll value
+
+    ; Store same value for all 12 scanlines of this row
+        ldy.w #0x000C  ; 12 scanlines
+
+    _reset_scanline_loop:
+        lda.b 0x00
+        sta.l HDMA_SWAP_Y, x  ; $81F4 + X
+        sta.l HDMA_ACTIVE_Y, x  ; $7F74 + X
+        inx
+        inx
+        inx
+        inx  ; X += 4
+        dey
+        bne _reset_scanline_loop
+
+    ; Next row
+        inc.b 0x02
+        lda.b 0x02
+        cmp.w #VISIBLE_ROWS  ; 5 rows
+        bne _reset_row_loop
+
+        sep #0x20  ; 8-bit A (game expects this)
+        ; Restore $00-$03 before returning
+        pla
+        sta.b 0x03
+        pla
+        sta.b 0x02
+        pla
+        sta.b 0x01
+        pla
+        sta.b 0x00
+        rtl
+
+    ; ============================================================================
+    ; check_cursor2_visibility_rolling
+    ; ============================================================================
+    ; Checks if cursor 2 (the first selected item in swap mode) should be visible.
+    ; Called after scroll animation completes.
+    ;
 
-;   - Else:
-;       Hide cursor 2 ($EF6A = 1)
-;
-; Called via JSR from wrap_and_clear_trampoline (bank $02)
+    ; Logic:
+    ;   - If swap mode NOT active ($EF94 == 0), return immediately
+    ;   - Get first selected item index from $EF95 (mask out bit 7)
 
-swap_mode_flag := 0xEF94  ; Non-zero = swap mode active
-first_selected_item := 0xEF95  ; First selected item index (bit 7 may be set)
-hide_cursor_2 := 0xEF6A  ; Non-zero = hide cursor 2
+    ;   - If rolling_top_row <= selected < rolling_top_row + VISIBLE_ROWS:
+    ;       Show cursor 2 ($EF6A = 0)
 
-check_cursor2_visibility_rolling:
-"""
-Toggle cursor-2 visibility while swap mode is active based on whether the first-selected item is in the visible
-window.
-"""
-; Check if we're in inventory mode (bit 2 of $4A)
-; If not, don't touch cursor 2 state
-    lda.b 0x4A
-    and #0x04
-    beq _cursor2_done
-
-; Check if swap mode is active
-    lda.w swap_mode_flag
-    beq _cursor2_done  ; Not in swap mode, nothing to check
-
-; Get first selected item index (mask out bit 7)
-    lda.w first_selected_item
-    and #0x7F  ; Clear bit 7
-    sta.b 0x00  ; Save selected item index
-
-; Check if selected item is in visible range:
-; visible if: rolling_top_row <= selected < rolling_top_row + VISIBLE_ROWS
-
-; First check: selected >= rolling_top_row
-    lda.b 0x00  ; Selected item
-    cmp.w rolling_top_row
-    bcc _cursor2_hide  ; selected < rolling_top_row, hide cursor
-
-; Second check: selected < rolling_top_row + VISIBLE_ROWS
-    lda.w rolling_top_row
-    clc
-    adc #VISIBLE_ROWS  ; rolling_top_row + 5
-    sta.b 0x01  ; Save upper bound
-
-    lda.b 0x00  ; Selected item
-    cmp.b 0x01  ; Compare to upper bound
-    bcs _cursor2_hide  ; selected >= upper bound, hide cursor
-
-; Item is visible - show cursor 2
-    stz.w hide_cursor_2  ; $EF6A = 0 (show)
-    bra _cursor2_done
-
-_cursor2_hide:
-    ; Item is not visible - hide cursor 2
-    lda #0x01
-    sta.w hide_cursor_2  ; $EF6A = 1 (hide)
-
-_cursor2_done:
-    rtl  ; Called via JSL from bank $02
-
-; ============================================================================
-; Field Menu NMI Handler (relocated from bank $01 to save space)
-; ============================================================================
-; Constants for field menu HDMA (duplicated here for bank $20 access)
-field_menu_hdma_enable := 0x1BAE
-field_menu_hdma_copy_pending := 0x1BB6
-field_menu_transfer_pending := 0x1BB3
-FIELD_HDMA_TABLE := 0x7E9800
-FIELD_HDMA_SHADOW := 0x7E9840
-FIELD_HDMA_TABLE_SIZE := 40
-
-; Drops profile HDMA constants (mirror src/ingame/drops_rolling.s
-; definitions; duplicated here so the NMI handler can reference them
-; without crossing the bank-$01 / bank-$20 import boundary).
-DROPS_HDMA_TABLE := 0x7E9880
-DROPS_HDMA_SHADOW := 0x7E98C0
-DROPS_HDMA_TABLE_SIZE := 40
-drops_hdma_copy_pending := 0x1BEE
-
-; Called via JSL from bank $01 nmi_dma_transfer_check
-
-field_menu_nmi_dma_transfer_check_impl:
-"""
-NMI hook (JSL from bank $01) relocated to bank $20: transfer the field rolling-buffer tilemap to VRAM and update
-the live HDMA scroll table.
-"""
-    php
-    sep #0x20  ; 8-bit A
-
-; === GUARD: Only run if menu HDMA is enabled ===
-; This prevents field menu DMA from corrupting battle VRAM
-    lda.l 0x7E0000 + field_menu_hdma_enable
-    bne _field_nmi_active
-    jmp.w _field_nmi_done
-
-_field_nmi_active:
-
-; === HDMA table copy: shadow -> active ===
-    lda.l 0x7E0000 + field_menu_hdma_copy_pending
-    beq _field_nmi_hdma_copy_done
-    lda #0x00
-    sta.l 0x7E0000 + field_menu_hdma_copy_pending
-
-; Copy 40 bytes from shadow ($9840) to active ($9800)
-    rep #0x30  ; 16-bit A, X, Y
-    ldx.w #0x0000
-
-_field_nmi_hdma_copy_loop:
-    lda.l FIELD_HDMA_SHADOW, x
-    sta.l FIELD_HDMA_TABLE, x
-    inx
-    inx
-    cpx.w #FIELD_HDMA_TABLE_SIZE
-    bcc _field_nmi_hdma_copy_loop
-    sep #0x20  ; Back to 8-bit A
-
-_field_nmi_hdma_copy_done:
-    ; === Drops HDMA table copy: $7E:98C0 shadow -> $7E:9880 active ===
-    ; Drops uses ch4 driving BG4VOFS with its own 64-byte table, so the
-    ; shadow→active copy needs a parallel block keyed off
-    ; drops_hdma_copy_pending. Skip if drops HDMA is disabled in
-    ; $1BAE bit 4.
-    sep #0x20
-    lda.l 0x7E1BAE
-    and #0x10
-    beq _drops_nmi_hdma_copy_done
-    lda.l 0x7E0000 + drops_hdma_copy_pending
-    beq _drops_nmi_hdma_copy_done
-    lda #0x00
-    sta.l 0x7E0000 + drops_hdma_copy_pending
-    rep #0x30
-    ldx.w #0x0000
-
-_drops_nmi_hdma_copy_loop:
-    lda.l DROPS_HDMA_SHADOW, x
-    sta.l DROPS_HDMA_TABLE, x
-    inx
-    inx
-    cpx.w #DROPS_HDMA_TABLE_SIZE
-    bcc _drops_nmi_hdma_copy_loop
-    sep #0x20
-
-_drops_nmi_hdma_copy_done:
-
-; === Tilemap DMA transfer (field menu = BG1) ===
-; Skip when treasure menu owns the screen: $1BB3 is then the
-; original drops cursor row, not field_menu_transfer_pending,
-; and clearing it would snap the drops cursor back to row 0.
-    lda.l 0x7E1BC6
-    bne _field_nmi_check_treasure
-    lda.l 0x7E0000 + field_menu_transfer_pending
-    beq _field_nmi_check_treasure
-    lda #0x00
-    sta.l 0x7E0000 + field_menu_transfer_pending
-
-    sep #0x20
-    lda #0x01
-    sta.w 0x4300
-    lda #0x18
-    sta.w 0x4301
-    rep #0x20
-    lda.w #0xB600
-    sta.w 0x4302
-    sep #0x20
-    lda #0x7E
-    sta.w 0x4304
-    rep #0x20
-    lda.w #0x0800
-    sta.w 0x4305
-    lda.w #0x6000
-    sta.w 0x2116
-    sep #0x20
-    lda #0x01
-    sta.w 0x420B
+    ;   - Else:
+    ;       Hide cursor 2 ($EF6A = 1)
+    ;
+    ; Called via JSR from wrap_and_clear_trampoline (bank $02)
 
-_field_nmi_check_treasure:
-.if TREASURE_INVENTORY_ROLLING {
-; === Tilemap DMA transfer (treasure menu = BG3) ===
-    lda.l 0x7E0000 + 0x1BDB  ; treasure_transfer_pending
-    beq _field_nmi_done
-    lda #0x00
-    sta.l 0x7E0000 + 0x1BDB
+    swap_mode_flag := 0xEF94  ; Non-zero = swap mode active
+    first_selected_item := 0xEF95  ; First selected item index (bit 7 may be set)
+    hide_cursor_2 := 0xEF6A  ; Non-zero = hide cursor 2
 
-    sep #0x20
-    lda #0x01
-    sta.w 0x4300
-    lda #0x18
-    sta.w 0x4301
-    rep #0x20
-    lda.w #0xD600  ; BG3 screen buffer source ($7ED600)
-    sta.w 0x4302
-    sep #0x20
-    lda #0x7E
-    sta.w 0x4304
-    rep #0x20
-    lda.w #0x0800  ; 2 KB tilemap
-    sta.w 0x4305
-    lda.w #0x7000  ; BG3 tilemap VRAM target
-    sta.w 0x2116
-    sep #0x20
-    lda #0x01
-    sta.w 0x420B
+    check_cursor2_visibility_rolling:
+    """
+    Toggle cursor-2 visibility while swap mode is active based on whether the first-selected item is in the visible
+    window.
+    """
+    ; Check if we're in inventory mode (bit 2 of $4A)
+    ; If not, don't touch cursor 2 state
+        lda.b 0x4A
+        and #0x04
+        beq _cursor2_done
+
+    ; Check if swap mode is active
+        lda.w swap_mode_flag
+        beq _cursor2_done  ; Not in swap mode, nothing to check
+
+    ; Get first selected item index (mask out bit 7)
+        lda.w first_selected_item
+        and #0x7F  ; Clear bit 7
+        sta.b 0x00  ; Save selected item index
+
+    ; Check if selected item is in visible range:
+    ; visible if: rolling_top_row <= selected < rolling_top_row + VISIBLE_ROWS
+
+    ; First check: selected >= rolling_top_row
+        lda.b 0x00  ; Selected item
+        cmp.w rolling_top_row
+        bcc _cursor2_hide  ; selected < rolling_top_row, hide cursor
+
+    ; Second check: selected < rolling_top_row + VISIBLE_ROWS
+        lda.w rolling_top_row
+        clc
+        adc #VISIBLE_ROWS  ; rolling_top_row + 5
+        sta.b 0x01  ; Save upper bound
+
+        lda.b 0x00  ; Selected item
+        cmp.b 0x01  ; Compare to upper bound
+        bcs _cursor2_hide  ; selected >= upper bound, hide cursor
+
+    ; Item is visible - show cursor 2
+        stz.w hide_cursor_2  ; $EF6A = 0 (show)
+        bra _cursor2_done
+
+    _cursor2_hide:
+        ; Item is not visible - hide cursor 2
+        lda #0x01
+        sta.w hide_cursor_2  ; $EF6A = 1 (hide)
+
+    _cursor2_done:
+        rtl  ; Called via JSL from bank $02
+
+    ; ============================================================================
+    ; Field Menu NMI Handler (relocated from bank $01 to save space)
+    ; ============================================================================
+    ; Constants for field menu HDMA (duplicated here for bank $20 access)
+    field_menu_hdma_enable := 0x1BAE
+    field_menu_hdma_copy_pending := 0x1BB6
+    field_menu_transfer_pending := 0x1BB3
+    FIELD_HDMA_TABLE := 0x7E9800
+    FIELD_HDMA_SHADOW := 0x7E9840
+    FIELD_HDMA_TABLE_SIZE := 40
+
+    ; Drops profile HDMA constants (mirror src/ingame/drops_rolling.s
+    ; definitions; duplicated here so the NMI handler can reference them
+    ; without crossing the bank-$01 / bank-$20 import boundary).
+    DROPS_HDMA_TABLE := 0x7E9880
+    DROPS_HDMA_SHADOW := 0x7E98C0
+    DROPS_HDMA_TABLE_SIZE := 40
+    drops_hdma_copy_pending := 0x9C3E  ; drops_rolling + 0x0E (drops_rolling moved to $9C30)
+
+    ; Called via JSL from bank $01 nmi_dma_transfer_check
+
+    field_menu_nmi_dma_transfer_check_impl:
+    """
+    NMI hook (JSL from bank $01) relocated to bank $20: transfer the field rolling-buffer tilemap to VRAM and update
+    the live HDMA scroll table.
+    """
+        php
+        sep #0x20  ; 8-bit A
+
+    ; === GUARD: Only run if menu HDMA is enabled ===
+    ; This prevents field menu DMA from corrupting battle VRAM
+        lda.l 0x7E0000 + field_menu_hdma_enable
+        bne _field_nmi_active
+        jmp.w _field_nmi_done
+
+    _field_nmi_active:
+
+    ; === HDMA table copy: shadow -> active ===
+        lda.l 0x7E0000 + field_menu_hdma_copy_pending
+        beq _field_nmi_hdma_copy_done
+        lda #0x00
+        sta.l 0x7E0000 + field_menu_hdma_copy_pending
+
+    ; Copy 40 bytes from shadow ($9840) to active ($9800)
+        rep #0x30  ; 16-bit A, X, Y
+        ldx.w #0x0000
+
+    _field_nmi_hdma_copy_loop:
+        lda.l FIELD_HDMA_SHADOW, x
+        sta.l FIELD_HDMA_TABLE, x
+        inx
+        inx
+        cpx.w #FIELD_HDMA_TABLE_SIZE
+        bcc _field_nmi_hdma_copy_loop
+        sep #0x20  ; Back to 8-bit A
+
+    _field_nmi_hdma_copy_done:
+        ; === Drops HDMA table copy: $7E:98C0 shadow -> $7E:9880 active ===
+        ; Drops uses ch4 driving BG4VOFS with its own 64-byte table, so the
+        ; shadow→active copy needs a parallel block keyed off
+        ; drops_hdma_copy_pending. Skip if drops HDMA is disabled in
+        ; $1BAE bit 4.
+        sep #0x20
+        lda.l 0x7E1BAE
+        and #0x10
+        beq _drops_nmi_hdma_copy_done
+        lda.l 0x7E0000 + drops_hdma_copy_pending
+        beq _drops_nmi_hdma_copy_done
+        lda #0x00
+        sta.l 0x7E0000 + drops_hdma_copy_pending
+        rep #0x30
+        ldx.w #0x0000
+
+    _drops_nmi_hdma_copy_loop:
+        lda.l DROPS_HDMA_SHADOW, x
+        sta.l DROPS_HDMA_TABLE, x
+        inx
+        inx
+        cpx.w #DROPS_HDMA_TABLE_SIZE
+        bcc _drops_nmi_hdma_copy_loop
+        sep #0x20
+
+    _drops_nmi_hdma_copy_done:
+
+    ; === Tilemap DMA transfer (field menu = BG1) ===
+    ; Skip when treasure menu owns the screen: $1BB3 is then the
+    ; original drops cursor row, not field_menu_transfer_pending,
+    ; and clearing it would snap the drops cursor back to row 0.
+        lda.l 0x7E1BC6
+        bne _field_nmi_check_treasure
+        lda.l 0x7E0000 + field_menu_transfer_pending
+        beq _field_nmi_check_treasure
+        lda #0x00
+        sta.l 0x7E0000 + field_menu_transfer_pending
+
+        sep #0x20
+        lda #0x01
+        sta.w 0x4300
+        lda #0x18
+        sta.w 0x4301
+        rep #0x20
+        lda.w #0xB600
+        sta.w 0x4302
+        sep #0x20
+        lda #0x7E
+        sta.w 0x4304
+        rep #0x20
+        lda.w #0x0800
+        sta.w 0x4305
+        lda.w #0x6000
+        sta.w 0x2116
+        sep #0x20
+        lda #0x01
+        sta.w 0x420B
+
+    _field_nmi_check_treasure:
+    .if TREASURE_INVENTORY_ROLLING {
+    ; === Tilemap DMA transfer (treasure menu = BG3) ===
+        lda.l 0x7E0000 + 0x9C0B  ; treasure_transfer_pending
+        beq _field_nmi_done
+        lda #0x00
+        sta.l 0x7E0000 + 0x9C0B
+
+        sep #0x20
+        lda #0x01
+        sta.w 0x4300
+        lda #0x18
+        sta.w 0x4301
+        rep #0x20
+        lda.w #0xD600  ; BG3 screen buffer source ($7ED600)
+        sta.w 0x4302
+        sep #0x20
+        lda #0x7E
+        sta.w 0x4304
+        rep #0x20
+        lda.w #0x0800  ; 2 KB tilemap
+        sta.w 0x4305
+        lda.w #0x7000  ; BG3 tilemap VRAM target
+        sta.w 0x2116
+        sep #0x20
+        lda #0x01
+        sta.w 0x420B
+    }
+
+    _field_nmi_done:
+        ; --- VWF CHR flush (DMA channel 6) ---
+        ; Hands off to `render.flush_chr_to_vram`. Gates on `VWF_CHR_DIRTY`
+        ; and reads VRAM dest + size from `VwfConfig`. Drives ch6, which
+        ; the FF4 DMA audit shows untouched by vanilla btlgfx / menu and
+        ; by every engine we ship (ch7 is the shared battle / libmz /
+        ; field-NMI-tilemap channel, ch3 was the WRAM mirror experiment).
+        ; `flush_chr_to_vram` lives in the same bank-20 reloc region as
+        ; this impl ; use a short JSR so the return stays balanced with
+        ; the RTS the engine routine ends with.
+        jsr.w render.flush_chr_to_vram
+        plp
+        rtl
 }
-
-_field_nmi_done:
-    plp
-    rtl
diff --git a/src/battle/inventory_rolling_patches.s b/src/battle/inventory_rolling_patches.s
index c275220..4ace658 100644
--- a/src/battle/inventory_rolling_patches.s
+++ b/src/battle/inventory_rolling_patches.s
@@ -16,6 +16,13 @@ calls, JML hooks for the scroll animation, surgical NOPs / RTS overrides).
 .extern battle_render.TILEMAP_PENDING_COMMANDS
 .extern battle_menu_dirty
 .extern CMD_DIRTY_BIT
+.if BATTLE_ITEMS_VWF {
+    .extern messages_vwf.init_inventory
+    .extern messages_vwf.init_inventory_for_current_slot
+    .extern messages_vwf.mirror_main_to_cmd
+    .extern messages_vwf.draw_inventory_text
+    .extern messages_vwf.deinit
+}
 
 ; ============================================================================
 ; Rolling Inventory Buffer - ROM Patches (Single Column)
@@ -62,26 +69,35 @@ calls, JML hooks for the scroll animation, surgical NOPs / RTS overrides).
 }
 
 .alloc bank02_trampolines_block in bank02_trampolines {
-
 draw_text_rolling_trampoline:
-"""Bank-$02 trampoline: forces VWF battle-flags off for the duration of original draw_text call."""
-    ; Save current battle flags to prevent VWF mode interference
+"""
+Bank-$02 trampoline around draw_text for inventory rendering. With
+BATTLE_ITEMS_VWF on, sets battle_flags = 0x02 so battle_display_char
+routes the put_char dispatch through messages_vwf.put_fixed_char_*
+for proportional rendering. Wraps the call with init_names / deinit
+so the VWF tile allocator and pending-DMA mask stay in sync.
+"""
+
+
+    .if BATTLE_ITEMS_VWF {
+; Custom draw_inventory_text owns the format walk end-to-end (escape
+; codes 0x00 / 0x03 / 0x0E plus VWF blits for raw chars). No need to
+; toggle battle_flags here ; it manages its own VWF state.
+    jsr.l messages_vwf.draw_inventory_text
+    rtl
+    } else {
     lda.l 0x704F00
     pha
-    ; Clear battle flags (force WRAM mode for inventory)
-    lda #0x00
+    lda.b #0x00
     sta.l 0x704F00
-
-; Original draw_text call
     xba
-    lda #0x00
+    lda.b #0x00
     xba
-    jsr 0xA455  ; draw_text at $02A455
-
-; Restore battle flags
+    jsr 0xA455
     pla
     sta.l 0x704F00
     rtl
+    }
 
 mult8_trampoline:
 """Bank-$02 RTL trampoline around original Mult8 ($028560)."""
@@ -113,26 +129,23 @@ _update_enabled_items_trampoline:
 ; Original function was overwritten. This must fit in bank $02 free space.
 
 _draw_battle_command_window_relocated:
-    ; Thunk-level gate on CMD_DIRTY_BIT. Set TILEMAP_PENDING_COMMANDS
-    ; so the NMI dma_transfer picks up the cmd-window tilemap DMA
-    ; in sync with the tile-data DMA. Skip on clean ; VRAM retains
-    ; last upload and no DMA fires.
-    lda.l battle_menu_dirty
-    bit.b #CMD_DIRTY_BIT
-    beq _dbcw_skip
-    and.b #( ~ CMD_DIRTY_BIT ) & 0xFF
-    sta.l battle_menu_dirty
+; Drop the CMD_DIRTY_BIT gate. The cmd-window tilemap region at
+; $C1A5+ is a mirror of the main view ($BE65+) overlaid with cmd
+; tiles. ATB rotation / monster death / HP ticks etc. only write to
+; main, never re-mirror, so gating left the cmd region frozen at
+; battle-init state (and we kept seeing empty char-name + monster
+; rows behind the cmd window). Mirror is now a single ch3 WRAM DMA
+; (~400 cycles), small enough to run every frame; total cost is in
+; the same ballpark as vanilla's per-frame DrawCmdWindow.
     lda.l battle_render.tilemap_pending_mask
     ora.b #battle_render.TILEMAP_PENDING_COMMANDS
     sta.l battle_render.tilemap_pending_mask
 
-    jsr.w draw_window_render_hook  ; Draw command list + sets LDX #$0340
+    jsr.w draw_window_render_hook  ; Draw command list (X side-effect unused now)
 
-_dbcw_loop:
-    lda.w 0xBE65, x  ; Copy main menu window tilemap
-    sta.w 0xC1A5, x
-    dex
-    bne _dbcw_loop
+; Mirror main-view tilemap $BE65..$C1A4 -> $C1A5..$C4E4 via WRAM DMA
+; ch3 (replaces a $340-iter lda/sta loop ; ~10K cycles -> ~400).
+    jsr.l messages_vwf.mirror_main_to_cmd
     lda #0x02
     jsr 0x9B59  ; Load menu window data
     jsr 0x9BC7  ; Draw window
@@ -140,9 +153,6 @@ _dbcw_loop:
     ldx.w #0x0064
     jmp 0x99F1  ; tail-call DrawCmdListText (rts via that function)
 
-_dbcw_skip:
-    rts
-
 ; ============================================================================
 ; WRAP/CLEAR TRAMPOLINE (small, stays in bank $02)
 ; ============================================================================
@@ -153,12 +163,12 @@ _dbcw_skip:
 
 wrap_and_clear_trampoline:
 """Bank-$02 tail of the scroll animation: post-render and cursor visibility check."""
-    ; FF6-style circular scroll - no reset needed!
-    ; The wrap function in update_list_scroll_hdma_wrapped handles coordinate conversion.
-    ;
-    ; After scroll down, we need to post-render the next item to prepare for
-    ; future scrolls. The off-screen slot that just scrolled off should be
-    ; updated with the next item in the list.
+; FF6-style circular scroll - no reset needed!
+; The wrap function in update_list_scroll_hdma_wrapped handles coordinate conversion.
+;
+; After scroll down, we need to post-render the next item to prepare for
+; future scrolls. The off-screen slot that just scrolled off should be
+; updated with the next item in the list.
     jsr.l post_scroll_down_render  ; Post-render if scroll down
 
 ; Check cursor 2 visibility for swap mode
@@ -174,8 +184,9 @@ wrap_and_clear_trampoline:
 return_to_bank02:
 """Trailing RTS used as a JML target by bank-$20 hooks to return to bank-$02."""
     rts
+}
 
-}  ; end .alloc bank02_trampolines_block
+; end .alloc bank02_trampolines_block
 
 ; ============================================================================
 ; ROM PATCHES
diff --git a/src/battle/magic_reloc.s b/src/battle/magic_reloc.s
index 5d2f6dd..7708e1a 100644
--- a/src/battle/magic_reloc.s
+++ b/src/battle/magic_reloc.s
@@ -1,4 +1,6 @@
 """
+.include "../bank20.i"
+
 Relocated battle spell-list renderer (`draw_magic_list_direct`) and per-magic-type pointer table
 (`magic_list_ptrs`).
 """
@@ -10,165 +12,168 @@ destination_buffer = 0xc530 - 4
 left_column_base = destination_buffer - 4
 right_column_base = destination_buffer + 18 - 2
 
-draw_magic_list_direct:
-"""Relocated battle spell-list renderer: walks the per-character spell-list pointer table."""
-{
-    spell_id = 0x03
-    spell_enabled_flag = 0x02
-    current_row_offset = 0x08
-    spell_counter = 0x06
-    current_spell_index = 0x0a
-    lda 0x00
-; character slot
-    asl
-    tax
-    rep #0x20
-    lda.l magic_list_ptrs, x
-; pointers to spell lists
-    clc
-    adc 0x06
-; add magic type offset
-    sta 0x00
-    stz.b current_row_offset
-; start at row 0 (0x0000)
-    sep #0x20
-    lda #0x18
-; 24 spells total (12 rows x 2 columns)
-    sta.b spell_counter
-; spell counter
-    lda #0x00
-; current spell index (0-23)
-    sta.b current_spell_index
 
-spell_loop:
-    phx
-    pha
-    rep #0x20
-    lda 0x00
-    tax
-    sep #0x20
-    lda.w 0x0000, x
-    and #0x80
-    sta.b spell_enabled_flag
-    lda.w 0x0001, x
-    sta.b spell_id
-    pla
-    plx
-; Fast column selection using precomputed addresses
-    lda.b current_spell_index
-; spell index
-    and #0x01
-; check if odd
-    beq left_column
-; Right column
-    rep #0x20
-    lda.w #right_column_base
-    clc
-    adc.b current_row_offset
-; add current row offset
-    sta 0x32
-    bra set_second_addr
+.alloc battle_magic_reloc_block in bank20_reloc {
+    draw_magic_list_direct:
+    """Relocated battle spell-list renderer: walks the per-character spell-list pointer table."""
+    {
+        spell_id = 0x03
+        spell_enabled_flag = 0x02
+        current_row_offset = 0x08
+        spell_counter = 0x06
+        current_spell_index = 0x0a
+        lda 0x00
+    ; character slot
+        asl
+        tax
+        rep #0x20
+        lda.l magic_list_ptrs, x
+    ; pointers to spell lists
+        clc
+        adc 0x06
+    ; add magic type offset
+        sta 0x00
+        stz.b current_row_offset
+    ; start at row 0 (0x0000)
+        sep #0x20
+        lda #0x18
+    ; 24 spells total (12 rows x 2 columns)
+        sta.b spell_counter
+    ; spell counter
+        lda #0x00
+    ; current spell index (0-23)
+        sta.b current_spell_index
 
-left_column:
-; Left column
-    rep #0x20
-    lda.w #left_column_base
-    clc
-    adc.b current_row_offset
-; add current row offset
-    sta 0x32
+    spell_loop:
+        phx
+        pha
+        rep #0x20
+        lda 0x00
+        tax
+        sep #0x20
+        lda.w 0x0000, x
+        and #0x80
+        sta.b spell_enabled_flag
+        lda.w 0x0001, x
+        sta.b spell_id
+        pla
+        plx
+    ; Fast column selection using precomputed addresses
+        lda.b current_spell_index
+    ; spell index
+        and #0x01
+    ; check if odd
+        beq left_column
+    ; Right column
+        rep #0x20
+        lda.w #right_column_base
+        clc
+        adc.b current_row_offset
+    ; add current row offset
+        sta 0x32
+        bra set_second_addr
 
-set_second_addr:
-    adc.w #0x0040
-    sta 0x34
-    sep #0x20
-    ldy.w #0x00
-; Y offset
-    lda #0x00
-; tile flags
-    sta 0x36
-    lda.b spell_enabled_flag
-    beq enabled_spell
-    lda #0x04
-    sta 0x36
+    left_column:
+    ; Left column
+        rep #0x20
+        lda.w #left_column_base
+        clc
+        adc.b current_row_offset
+    ; add current row offset
+        sta 0x32
 
-enabled_spell:
-; Get spell ID and load spell name
-    lda.b spell_id
-; read spell ID
-    rep #0x20
-    and.w #0x007f
-; clear disabled bit
-    sep #0x20
-    sta.l 0x004202
-    lda.b #battle_magic_length
-    sta.l 0x004203
-    nop
-    nop
-    nop
-    nop
-    rep #0x20
-    lda.l 0x004216
-;     asl                 ; spell ID * 8 (8 bytes per name)
-;     asl
-;     asl
-    tax
-    sep #0x20
-    lda.b #battle_magic_length
-; 8 characters to write
-    sta 0x02
-; draw a space before rendering the magic name to clear left overs from the items.
-    lda #0xff
-    jsr.l draw_letter_far
+    set_second_addr:
+        adc.w #0x0040
+        sta 0x34
+        sep #0x20
+        ldy.w #0x00
+    ; Y offset
+        lda #0x00
+    ; tile flags
+        sta 0x36
+        lda.b spell_enabled_flag
+        beq enabled_spell
+        lda #0x04
+        sta 0x36
 
+    enabled_spell:
+    ; Get spell ID and load spell name
+        lda.b spell_id
+    ; read spell ID
+        rep #0x20
+        and.w #0x007f
+    ; clear disabled bit
+        sep #0x20
+        sta.l 0x004202
+        lda.b #battle_magic_length
+        sta.l 0x004203
+        nop
+        nop
+        nop
+        nop
+        rep #0x20
+        lda.l 0x004216
+    ;     asl                 ; spell ID * 8 (8 bytes per name)
+    ;     asl
+    ;     asl
+        tax
+        sep #0x20
+        lda.b #battle_magic_length
+    ; 8 characters to write
+        sta 0x02
+    ; draw a space before rendering the magic name to clear left overs from the items.
+        lda #0xff
+        jsr.l draw_letter_far
 
-letter_loop:
-    lda.l assets_magic_dat, x
-    jsr.l draw_letter_far
-    inx
-    dec 0x02
-    bne letter_loop
-    lda #0x0A
-    sta 0x02
 
-clear_loop:
-    lda.b #0xff
-    jsr.l draw_letter_far
-    dec 0x02
-    bne clear_loop
+    letter_loop:
+        lda.l assets_magic_dat, x
+        jsr.l draw_letter_far
+        inx
+        dec 0x02
+        bne letter_loop
+        lda #0x0A
+        sta 0x02
 
-next_spell:
-; Advance pointer by 4 bytes like original
-    rep #0x20
-    lda 0x00
-    clc
-    adc.w #0x0004
-    sta 0x00
-; Increment row offset after right column (odd spell index)
-    lda.b current_spell_index
-; current spell index
-    and.w #0x0001
-; check if odd (right column)
-    beq same_row
-; Just finished right column - move to next row
-    lda.b current_row_offset
-    clc
-    adc.w #0x0080
-    sta.b current_row_offset
+    clear_loop:
+        lda.b #0xff
+        jsr.l draw_letter_far
+        dec 0x02
+        bne clear_loop
 
-same_row:
-    sep #0x20
-    inc 0x0a
-; next spell index
-    dec.b spell_counter
-; decrement spell counter
-    beq exit
-    jmp.w spell_loop
+    next_spell:
+    ; Advance pointer by 4 bytes like original
+        rep #0x20
+        lda 0x00
+        clc
+        adc.w #0x0004
+        sta 0x00
+    ; Increment row offset after right column (odd spell index)
+        lda.b current_spell_index
+    ; current spell index
+        and.w #0x0001
+    ; check if odd (right column)
+        beq same_row
+    ; Just finished right column - move to next row
+        lda.b current_row_offset
+        clc
+        adc.w #0x0080
+        sta.b current_row_offset
 
-exit:
-    rtl
-}
+    same_row:
+        sep #0x20
+        inc 0x0a
+    ; next spell index
+        dec.b spell_counter
+    ; decrement spell counter
+        beq exit
+        jmp.w spell_loop
 
-magic_list_ptrs:
-"""Per-magic-type spell-list base pointers (white, black, summon, ninja, kokan)."""
-    .dw 0x2c7a, 0x2d9a, 0x2eba, 0x2fda, 0x30fa
+    exit:
+        rtl
+    }
+
+    magic_list_ptrs:
+    """Per-magic-type spell-list base pointers (white, black, summon, ninja, kokan)."""
+        .dw 0x2c7a, 0x2d9a, 0x2eba, 0x2fda, 0x30fa
+}
diff --git a/src/battle/math_reloc.s b/src/battle/math_reloc.s
index ca15d01..615abe4 100644
--- a/src/battle/math_reloc.s
+++ b/src/battle/math_reloc.s
@@ -1,86 +1,91 @@
 """
+.include "../bank20.i"
+
 Relocated 16x16 -> 32 multiply (`_hw_mult16`) using the SNES hardware multiplier,
 plus shorta / shorta0 size-toggle helpers.
 Called from JMP trampoline at $83B9 (in `math_patches.s`).
 Input:  $393D (16-bit) * $393F (16-bit). Output: $3941 (low 16-bit), $3943 (high 16-bit).
 """
 
-.macro shorta() {
-    """Switch A to 8-bit (`SEP #$20`)."""
-    sep #0x20
-}
 
-.macro shorta0() {
-    """Clear A and switch to 8-bit (TDC then `SEP #$20`)."""
-    tdc
-    shorta()
-}
+.alloc battle_math_reloc_block in bank20_reloc {
+    .macro shorta() {
+        """Switch A to 8-bit (`SEP #$20`)."""
+        sep #0x20
+    }
 
-_hw_mult16:
-    php
-    rep #0x30  ; 16-bit A, X, Y
-    stz 0x3941  ; Clear result low
-    stz 0x3943  ; Clear result high
-    sep #0x20  ; 8-bit A for hardware multiply registers
+    .macro shorta0() {
+        """Clear A and switch to 8-bit (TDC then `SEP #$20`)."""
+        tdc
+        shorta()
+    }
 
-; The SNES hardware multiplier only does 8x8->16
-; For 16x16->32, we need: (AH*256 + AL) * (BH*256 + BL)
-;   = AH*BH*65536 + AH*BL*256 + AL*BH*256 + AL*BL
-; We compute 4 partial products and add them with proper alignment.
+    _hw_mult16:
+        php
+        rep #0x30  ; 16-bit A, X, Y
+        stz 0x3941  ; Clear result low
+        stz 0x3943  ; Clear result high
+        sep #0x20  ; 8-bit A for hardware multiply registers
 
-; 1. Low(A) * Low(B) -> result bytes 0-1
-    lda 0x393d  ; A low byte
-    sta 0x4202  ; WRMPYA - multiplicand
-    lda 0x393f  ; B low byte
-    sta 0x4203  ; WRMPYB - multiplier (triggers multiply)
-    nop
-    nop  ; Wait 8 cycles for result
-    rep #0x20
-    lda 0x4216  ; RDMPYL/H - 16-bit result
-    sta 0x3941  ; Store in result bytes 0-1
+    ; The SNES hardware multiplier only does 8x8->16
+    ; For 16x16->32, we need: (AH*256 + AL) * (BH*256 + BL)
+    ;   = AH*BH*65536 + AH*BL*256 + AL*BH*256 + AL*BL
+    ; We compute 4 partial products and add them with proper alignment.
 
-; 2. Low(A) * High(B) -> add to result bytes 1-2
-    sep #0x20
-    lda 0x393d  ; A low byte
-    sta 0x4202
-    lda 0x3940  ; B high byte ($393F + 1)
-    sta 0x4203
-    nop
-    nop
-    rep #0x20
-    lda 0x4216
-    clc
-    adc 0x3942  ; Add to result bytes 1-2 (with carry to byte 3)
-    sta 0x3942
+    ; 1. Low(A) * Low(B) -> result bytes 0-1
+        lda 0x393d  ; A low byte
+        sta 0x4202  ; WRMPYA - multiplicand
+        lda 0x393f  ; B low byte
+        sta 0x4203  ; WRMPYB - multiplier (triggers multiply)
+        nop
+        nop  ; Wait 8 cycles for result
+        rep #0x20
+        lda 0x4216  ; RDMPYL/H - 16-bit result
+        sta 0x3941  ; Store in result bytes 0-1
 
-; 3. High(A) * Low(B) -> add to result bytes 1-2
-    sep #0x20
-    lda 0x393e  ; A high byte ($393D + 1)
-    sta 0x4202
-    lda 0x393f  ; B low byte
-    sta 0x4203
-    nop
-    nop
-    rep #0x20
-    lda 0x4216
-    clc
-    adc 0x3942  ; Add to result bytes 1-2
-    sta 0x3942
+    ; 2. Low(A) * High(B) -> add to result bytes 1-2
+        sep #0x20
+        lda 0x393d  ; A low byte
+        sta 0x4202
+        lda 0x3940  ; B high byte ($393F + 1)
+        sta 0x4203
+        nop
+        nop
+        rep #0x20
+        lda 0x4216
+        clc
+        adc 0x3942  ; Add to result bytes 1-2 (with carry to byte 3)
+        sta 0x3942
 
-; 4. High(A) * High(B) -> add to result bytes 2-3
-    sep #0x20
-    lda 0x393e  ; A high byte
-    sta 0x4202
-    lda 0x3940  ; B high byte
-    sta 0x4203
-    nop
-    nop
-    rep #0x20
-    lda 0x4216
-    clc
-    adc 0x3943  ; Add to result bytes 2-3
-    sta 0x3943
+    ; 3. High(A) * Low(B) -> add to result bytes 1-2
+        sep #0x20
+        lda 0x393e  ; A high byte ($393D + 1)
+        sta 0x4202
+        lda 0x393f  ; B low byte
+        sta 0x4203
+        nop
+        nop
+        rep #0x20
+        lda 0x4216
+        clc
+        adc 0x3942  ; Add to result bytes 1-2
+        sta 0x3942
 
-    shorta0()  ; TDC + SEP #$20 - clear high byte of A, 8-bit mode
-    plp
-    rtl
+    ; 4. High(A) * High(B) -> add to result bytes 2-3
+        sep #0x20
+        lda 0x393e  ; A high byte
+        sta 0x4202
+        lda 0x3940  ; B high byte
+        sta 0x4203
+        nop
+        nop
+        rep #0x20
+        lda 0x4216
+        clc
+        adc 0x3943  ; Add to result bytes 2-3
+        sta 0x3943
+
+        shorta0()  ; TDC + SEP #$20 - clear high byte of A, 8-bit mode
+        plp
+        rtl
+}
diff --git a/src/battle/message.s b/src/battle/message.s
index 3eb6003..199340d 100644
--- a/src/battle/message.s
+++ b/src/battle/message.s
@@ -2,6 +2,8 @@
 Battle-message tile renderer + VWF parser scopes (`battle_render` low-level blitter, `messages_vwf` high-level
 dialog-stream consumer).
 """
+.include "src/battle/inventory_budget.i"
+.include "src/vwf_state.i"
 .if 0 {
     .scope _vwf_tile_ring {
 ; Ring buffer for VWF tile allocation
@@ -148,13 +150,22 @@ _not_found:
     0x80 -> 0xB0 char names
     0xB0 -> 0xF0 commands ? this one is untested.
     """
-    buffer_ptr = 0x703000
+    buffer_ptr = VWF_CHR_BUFFER
     buffer_size = 8 * ( 128 + 32 ) * 2
     region_size = 48
-    pending_transfer_mask = 0x703c00
+; Gate state moved past the inventory tile slice ($703C00..$703EF0)
+; so the rolling pre-render does not stomp these bytes.
+    pending_transfer_mask = 0x703f00
+; Per-slot CHR dirty bitmask for the inventory rolling buffer. Bit N
+; set when slot N's CHR slice at $703000 + (slot_base + N*10)*16 has
+; been touched and needs a VRAM flush. NMI `dma_transfer` consumes
+; one or more bits per frame and DMAs only the dirty slot's 160-byte
+; slice instead of the full 4KB CHR region, fits in vblank without
+; forced blank.
+    dma_dirty_slots = 0x703f10
 ; --- Per-region dirty bits (normal sense: 1 = dirty, 0 = clean) ---
 ; Sits next to the DMA queue byte; writers SET bits on state change.
-    region_dirty_bits = 0x703c01
+    region_dirty_bits = 0x703f01
     REGION_DIRTY_MESSAGES = 0x01
     REGION_DIRTY_MONSTERS = 0x02
     REGION_DIRTY_NAMES = 0x04
@@ -163,14 +174,14 @@ _not_found:
 ; because the region was clean; $00 if it ran the full init.
 ; Used by deinit_with_gate to decide whether to signal DMA, and
 ; by the gated trampoline to decide whether to skip DrawText.
-    render_skipped = 0x703c02
+    render_skipped = 0x703f02
 ; Per-region tilemap-DMA pending bitmask. Set by `init_*_gated`
 ; on the render path  ; consumed by `dma_transfer` in NMI to fire
 ; a per-region tilemap DMA (WRAM tilemap -> BG VRAM). Decouples
 ; tilemap upload from the vanilla `TfrCmdWindow` / `TfrMainMenu`
 ; periodic queue so the tile-data + tilemap transfers stay in
 ; sync on the same NMI as the render.
-    tilemap_pending_mask = 0x703c03
+    tilemap_pending_mask = 0x703f03
     TILEMAP_PENDING_COMMANDS = 0x01
     TILEMAP_PENDING_MAIN = 0x02
     bits_left_on_tile = 0xA9
@@ -191,6 +202,37 @@ init_names:
 init_commands_list:
 """Initialize the renderer targeting the commands list region."""
     lda.b #region_size * 3
+    bra _init
+init_inventory_region:
+"""
+Reset the allocator to the inventory tile_id base (0xC0) once per
+rolling render pass. Subsequent per-item DrawText calls let the
+allocator increment naturally, so each item owns its own tile range
+(item N uses 0xC0 + N * width_in_tiles). Skips clear_buffer for the
+reasons noted on the other inventory entry — tile_id 0xC0+ would
+overrun the shared buffer into the state words at $703C00+.
+"""
+
+
+    lda.b #region_size * 4
+    sta.l pending_transfer_mask
+    jsr.w render_allocator.init_with_tile_id
+    .if ENABLE_KERNING_MENU {
+    stz.b prev_char
+    }
+    lda.b #0x08
+    sta.b bits_left_on_tile
+    stz.b temp
+    stz.b counter
+    rts
+render_allocator_init_with_tile_id_thunk:
+"""
+In-scope thunk so messages_vwf can reach the allocator via jsr.w
+across scope boundaries. A on entry = tile_id base.
+"""
+
+
+    jmp.w render_allocator.init_with_tile_id
 _init:
     sta.l pending_transfer_mask
     jsr.w render_allocator.init_with_tile_id
@@ -752,15 +794,28 @@ battle_msg_kerning_binary_ext:
     rtl
     }
 tilemap_write_no_inc:
-    lda.l render_allocator.allocated_tile_id
+"""
+Write a 10-bit tile_id reference at tilemap_offset.
+
+Low 8 bits of tile_id go in the entry's low byte  ; bits 8-9 ride
+the entry's high byte alongside the palette / flip / priority bits.
+With the 16-bit allocator, the entry's high byte is just
+(palette | allocator_high_byte) — no hardcoded +0x100 shift, so
+inventory tile_ids past 0xFF reach the upper half of BG3 CHR
+cleanly instead of wrapping back into the messages region.
+"""
+
+
     phy
     ldy.b tilemap_offset
+    lda.l render_allocator.allocated_tile_id
     sta (0x34), y
     lda #0xff
     sta (0x32), y
     iny
     lda 0x36
     sta (0x32), y
+    ora.l render_allocator.allocated_tile_id + 1
     ora.b #0x01
     sta (0x34), y
     ply
@@ -829,6 +884,391 @@ init_names:
     jsr.l battle_flags.set_vwf_render
     jsr.w battle_render.init_names
     rtl
+init_inventory_for_current_slot:
+"""
+Compute tile_id base from the live rolling_slot_index (0..5),
+allocate 9 tile_ids per slot starting at 0xC0, and reset allocator
+state. Saves the bank-02 trampoline ~20 bytes by keeping the math
+here.
+"""
+
+
+    php
+    rep #0x10
+    rep #0x20
+    lda.l 0x7EEF82
+    and.w #0x00FF
+    sta.b 0x00
+
+; Tile_id base = slot_index * ITEM_VWF_TILE_BUDGET + ITEM_VWF_TILE_BASE.
+; With K=10 and BUFFER_SLOTS=6 that's 60 tiles in $C0..$FB, leaving 4
+; tiles of slack before the 8-bit wrap point. Math below assumes K=10
+; (asl x3 = *8, then +slot twice = *10) ; bump together if K changes.
+    asl
+    asl
+    asl
+    clc
+    adc.b 0x00
+    clc
+    adc.b 0x00
+    clc
+    adc.w #ITEM_VWF_TILE_BASE
+    sta.b 0x02
+
+; Clear the slot's CHR slice (ITEM_VWF_CHR_BYTES bytes) at
+; buffer_ptr + tile_id_base * 16. Stale bits would OR into the new
+; render via the VWF blitter's ora-then-store path. Pattern $FF $00
+; alternates the 2 bitplanes (empty tile in 4bpp). Loop count =
+; ITEM_VWF_CHR_WORDS = K * 8.
+    lda.b 0x02
+    asl
+    asl
+    asl
+    asl
+    clc
+    adc.w #battle_render.buffer_ptr & 0xffff
+    tax
+    ldy.w #ITEM_VWF_CHR_WORDS
+_chr_clear_loop:
+    lda.w #0x00ff
+    sta.l 0x700000, x
+    inx
+    inx
+    dey
+    bne _chr_clear_loop
+
+; Restore tile_id base into A for the allocator init.
+    lda.b 0x02
+    sep #0x20
+    pha
+    jsr.l battle_flags.set_vwf_render
+    pla
+    sta.l battle_render.pending_transfer_mask
+    jsr.w battle_render.render_allocator_init_with_tile_id_thunk
+; Slot owns ITEM_VWF_TILE_BUDGET tile_ids. Set the allocator clamp at
+; slot_base + (K-1) AFTER init_with_tile_id (which resets slot_limit_low
+; to 0xFF). Overflow freezes at the last tile instead of bleeding into
+; the next slot's CHR / tile_id range.
+    lda.b 0x02
+    clc
+    adc.b #( ITEM_VWF_TILE_BUDGET - 1 )
+    sta.l render_allocator.slot_limit_low
+    .if ENABLE_KERNING_MENU {
+    stz.b battle_render.prev_char
+    }
+    lda.b #0x08
+    sta.b battle_render.bits_left_on_tile
+    stz.b battle_render.temp
+    stz.b battle_render.counter
+    plp
+    rtl
+init_inventory:
+"""
+Wrap battle_render.init_inventory_region with set_vwf_render so the
+message renderer routes char dispatch through the VWF put_char path.
+"""
+
+
+    jsr.l battle_flags.set_vwf_render
+    jsr.w battle_render.init_inventory_region
+    rtl
+mirror_main_to_cmd:
+"""
+Block-mirror the main-view BG3 tilemap region $7E:BE65..$C1A4
+($340 bytes covering entry-0 + window frame) onto the cmd-window
+region $7E:C1A5..$C4E4 via the 65816 `MVN` block move. ~7 cycles
+per byte  ; ~5800 cycles total, roughly half the cost of the
+original $340-iter lda/sta loop that this replaces.
+
+First pass tried WRAM->WRAM GP-DMA via $2180 (WMDATA). It writes
+zeros: the A-bus read of source WRAM contends with the B-bus
+write into WRAM through $2180, the WRAM controller can not serve
+both halves of the cycle, and the destination ends up cleared.
+MVN routes bytes through the CPU one at a time so the bus stays
+single-master and the move actually lands.
+
+Caller assumed P with M=8 X=16 on entry, restored on exit.
+"""
+
+
+    php
+    rep #0x30
+    ldx.w #0xBE65
+    ldy.w #0xC1A5
+    lda.w #0x033F
+; A = byte count - 1 ($340 bytes)
+    .db 0x54
+    .db 0x7E
+    .db 0x7E
+; mvn #$7E, #$7E (dst_bank, src_bank)
+    plp
+    rtl
+draw_inventory_text:
+"""
+Custom inventory text renderer that walks a format buffer ourselves
+instead of going through the vanilla draw_text dispatch. Owns the
+escape interpretation end-to-end so we can mix VWF-rendered name
+chars with fixed-width symbol / colon / digit / type tiles without
+toggling battle_flags mid-stream.
+
+Caller setup mirrors the vanilla draw_text contract:
+  $EF50: 16-bit format buffer ptr (source bytes)
+  $EF52: 16-bit destination tilemap-buffer ptr (in bank \$7E)
+  $EF54: line length (tiles per row, currently 15)
+  $EF55: palette
+  $7EEF82: rolling_slot_index (0..5) for VWF allocator base
+
+Escape codes handled:
+  0x00: terminator -> return
+  0x03 BB: emit fixed tile_id BB at the next tilemap slot
+  0x0E PP: set tile-attribute byte to PP for subsequent writes
+  any other byte: VWF blit through battle_render.display_char
+"""
+
+
+    php
+    rep #0x10
+    sep #0x20
+
+; Set DBR to $7E so 16-bit absolute,Y / abs,X writes target WRAM.
+    phb
+    lda.b #0x7E
+    pha
+    plb
+
+; Init VWF state via the slot helper. It resets the allocator to
+; (0xC0 + slot * 10), bits_left_on_tile = 8, etc.
+    jsr.w init_inventory_for_current_slot_local
+
+; Vanilla draw_text prologue: stash src ptr / dest row-1 ptr / dest
+; row-2 ptr / palette into DP $30 / $32 / $34 / $36. tilemap_write
+; relies on these for the indirect ($32),y and ($34),y writes.
+    lda.w 0xef55
+    sta.b 0x36
+    ldx.w 0xef50
+    stx.b 0x30
+    ldx.w 0xef52
+    stx.b 0x32
+; Row 2 ptr = row 1 ptr + (line_length * 2). DON'T modify $ef54 in
+; place ; the caller hands us the live line_length each call, and a
+; persistent asl would double it every render.
+    lda.w 0xef54
+    asl
+    clc
+    adc.b 0x32
+    sta.b 0x34
+    lda.b 0x33
+    adc #0x00
+    sta.b 0x35
+
+; Pre-clear both row buffers in the slot with space tiles ($FF) +
+; palette so an `0xFC NN` goto can skip ahead without leaving stale
+; tile_ids on the way. 15 tiles per row, 2 bytes per entry = 30 byte
+; pairs per row.
+    ldy.w #0x0000
+_di_clear:
+    lda #0xff
+    sta (0x32), y
+    sta (0x34), y
+    iny
+    lda.b 0x36
+    sta (0x32), y
+    sta (0x34), y
+    iny
+    cpy.w #30
+    bne _di_clear
+
+; X = src (format buffer ptr), Y = dest offset (0-relative into the
+; tilemap buffer pointed to by \$32 / \$34).
+    ldx.w 0xEF50
+    ldy.w #0x0000
+
+; Control codes:
+;   0x00         -> terminate
+;   0x03 BB      -> fixed tile_id BB at current pos
+;   0x0E PP      -> set palette / tile-flag byte
+;   0xFC NN      -> goto tile slot NN in the slot tilemap (no fill ;
+;                  init pre-clears the slot to space tiles)
+;   0xFE         -> toggle fixed <-> VWF dispatch for subsequent chars
+; Default mode at entry = fixed so the leading symbol byte lands as a
+; literal tile_id without an explicit escape. Stash the mode flag in
+; scratch DP $1F: $37/$38/$39/$3A are controller-1/2 button shadows
+; and writing $80 to $37 set the A-bit, which cursor.asm @b5a6 then
+; latched as a phantom A press, toggling swap mode every scroll.
+    lda.b #0x80
+    sta.b 0x1F
+
+_di_loop:
+    lda.w 0x0000, x
+    beq _di_done
+    cmp #0x03
+    beq _di_fixed
+    cmp #0x0E
+    beq _di_pal
+    cmp #0xFC
+    beq _di_pad
+    cmp #0xFE
+    beq _di_vwf_toggle
+    bit.b 0x1F
+    bmi _di_fixed_char
+; VWF char -> battle_render.display_char (uses Y as tilemap_offset,
+; advances both source X and dest Y as it allocates).
+    inx
+    sty.b battle_render.tilemap_offset
+    jsr.w battle_render.display_char
+    ldy.b battle_render.tilemap_offset
+    jmp.w _di_loop
+
+_di_fixed_char:
+; Fixed-mode raw char: write current byte as tile_id at (\$34),y. No
+; ora #0x01 on the attr -- the +0x100 high bit is only for VWF tiles
+; in the 0x1xx range, fixed font tiles live in 0x00..0xFF.
+    sta (0x34), y
+    lda #0xff
+    sta (0x32), y
+    iny
+    lda.b 0x36
+    sta (0x32), y
+    sta (0x34), y
+    iny
+    inx
+    jmp.w _di_loop
+
+
+_di_fixed:
+; 0x03 BB -> write fixed tile_id BB at (\$34),y. Same shape as
+; _di_fixed_char (mirror of wram.put_char), no +0x100 bit set.
+    inx
+    lda.w 0x0000, x
+    sta (0x34), y
+    lda #0xff
+    sta (0x32), y
+    iny
+    lda.b 0x36
+    sta (0x32), y
+    sta (0x34), y
+    iny
+    inx
+    jmp.w _di_loop
+
+_di_pal:
+; 0x0E PP -> stash PP as the tile flags byte. Vanilla draw_text
+; stores it at $36 ; mirror that so subsequent fixed / VWF writes
+; pick up the right palette.
+    inx
+    lda.w 0x0000, x
+    sta.b 0x36
+    inx
+    jmp.w _di_loop
+
+_di_done:
+; Clear the VWF battle_flag so later non-inventory renders (monster HP
+; refresh, status text, etc.) go back to the WRAM put_char path.
+    jsr.l battle_flags.clear_vwf_render
+; Mark this slot's CHR slice dirty in dma_dirty_slots. NMI
+; dma_transfer picks it up and DMAs the 160-byte slice to VRAM.
+; Replaces the old full-4KB DMA trigger ; partial DMA fits in
+; vblank without forced blank.
+    php
+    sep #0x30
+    lda.l 0x7EEF82  ; slot index (low byte only ; M=8)
+    and.b #0x07  ; clamp to 0..7 (valid range 0..5)
+    tax
+    lda.l _dma_slot_bit_lut, x
+    ora.l battle_render.dma_dirty_slots
+    sta.l battle_render.dma_dirty_slots
+    plp
+    plb
+    plp
+    rtl
+
+_di_skip:
+    inx
+    jmp.w _di_loop
+
+_di_vwf_toggle:
+    inx
+    lda.b 0x1F
+    eor.b #0x80
+    sta.b 0x1F
+    jmp.w _di_loop
+
+_di_pad:
+; 0xFC NN -> goto tile slot NN (jump Y to NN * 2). The slot tilemap is
+; pre-cleared to space tiles in init, so intermediate slots already
+; show as blank without an explicit fill. Also resets bits_left so the
+; next VWF char starts on a clean tile column.
+    inx
+    lda.w 0x0000, x
+    inx
+    asl
+    tay
+    lda #0x08
+    sta.b battle_render.bits_left_on_tile
+    jmp.w _di_loop
+
+init_inventory_for_current_slot_local:
+"""
+Local jsr.w-reachable copy of init_inventory_for_current_slot that
+returns via rts (the public entry returns rtl).
+"""
+
+
+    php
+    rep #0x10
+    rep #0x20
+    lda.l 0x7EEF82
+    and.w #0x00FF
+    sta.b 0x00
+    asl
+    asl
+    asl
+    clc
+    adc.b 0x00
+    clc
+    adc.b 0x00
+    clc
+    adc.w #ITEM_VWF_TILE_BASE
+    sta.b 0x02
+    asl
+    asl
+    asl
+    asl
+    clc
+    adc.w #battle_render.buffer_ptr & 0xffff
+    tax
+    ldy.w #ITEM_VWF_CHR_WORDS
+_dis_chr_clear:
+    lda.w #0x00ff
+    sta.l 0x700000, x
+    inx
+    inx
+    dey
+    bne _dis_chr_clear
+    lda.b 0x02
+    sep #0x20
+    pha
+    jsr.l battle_flags.set_vwf_render
+    pla
+    sta.l battle_render.pending_transfer_mask
+    jsr.w battle_render.render_allocator_init_with_tile_id_thunk
+; Slot owns ITEM_VWF_TILE_BUDGET tile_ids. Set the allocator clamp at
+; slot_base + (K-1) AFTER init_with_tile_id (which resets slot_limit_low
+; to 0xFF). Overflow freezes at the last tile instead of bleeding into
+; the next slot's CHR / tile_id range.
+    lda.b 0x02
+    clc
+    adc.b #( ITEM_VWF_TILE_BUDGET - 1 )
+    sta.l render_allocator.slot_limit_low
+    .if ENABLE_KERNING_MENU {
+    stz.b battle_render.prev_char
+    }
+    lda.b #0x08
+    sta.b battle_render.bits_left_on_tile
+    stz.b battle_render.temp
+    stz.b battle_render.counter
+    plp
+    rts
 init_monsters_gated:
 """
 Gated counterpart to `init_monsters`. Always flips the VWF flag
@@ -907,6 +1347,120 @@ normal length, no visible black strip.
     pha
     phx
     phy
+    .if BATTLE_ITEMS_VWF {
+; Per-NMI BG3 V-scroll footer override.
+;
+; Channel 2 (BG3 V-scroll, indirect, $7E:760B -> $7E:7ED2 chunk) is
+; multiplexed across the BG3 layer (cmd / status / inventory share).
+; Vanilla 0x02BB at $7E:8068..$8081 leaves 2 rows of black below
+; inv body where the window's bottom border should be.
+;
+; While inv is open: write border-row scrolls every NMI (vanilla
+; state-1 swap at $02:A8FE would otherwise overwrite within a frame).
+; On the close edge: one-shot restore 0x02BB so cmd/status menus
+; that share ch2 stop rendering inventory's border tiles.
+;
+; Open-state shadow at $703F04 (next free past tilemap_pending_mask).
+    php
+    sep #0x30
+    lda.l 0x7E004A
+    and.b #0x04
+    beq _inv_footer_closed
+; OPEN: pin the two bottom-border tile rows past body. $8068+ is
+; past ch2's HDMA chunk (240-byte count terminates at $7FC2) ;
+; these writes don't drive HDMA but mirror to the swap table via
+; the vanilla state-1 routine which the inventory code reads. Real
+; mid-scroll leak fix needs to patch INSIDE chunk 2 ($7F74-$7FBF)
+; where vanilla rewrites body-row scrolls per rolling_buffer_pos.
+    rep #0x30
+    lda.w #0x0193
+    sta.l 0x7E8068
+    sta.l 0x7E806C
+    sta.l 0x7E8070
+    lda.w #0x019B
+    sta.l 0x7E8074
+    sta.l 0x7E8078
+    sta.l 0x7E807C
+    sta.l 0x7E8080
+    bra _inv_footer_done
+_inv_footer_closed:
+; CLOSED: write vanilla idle pattern every frame so state-1 swap
+; can't bring our open values back. Trace at battle-settle:
+;   $8064 = 0x0187 (vanilla "slot 5" transition entry, static)
+;   $8068/6C/70 = 0x01F7
+;   $8074..$80 = 0x0026, 0x0025, 0x0024, 0x0023 (decreasing)
+    rep #0x30
+    lda.w #0x01F7
+    sta.l 0x7E8068
+    sta.l 0x7E806C
+    sta.l 0x7E8070
+    lda.w #0x0026
+    sta.l 0x7E8074
+    lda.w #0x0025
+    sta.l 0x7E8078
+    lda.w #0x0024
+    sta.l 0x7E807C
+    lda.w #0x0023
+    sta.l 0x7E8080
+_inv_footer_done:
+    plp
+; --- Inventory CHR partial DMA ---
+; Pop one bit from dma_dirty_slots and DMA that slot's 160-byte CHR
+; slice ($703000 + (0xC0 + N*10)*16 -> VRAM $BC00 + N*$A0). Replaces
+; the old full-4KB DMA path for inventory. Non-inv VWF regions still
+; flow through the legacy pending_transfer_mask check below.
+    php
+    sep #0x20
+    rep #0x10
+    lda.l battle_render.dma_dirty_slots
+    and.b #0x3F
+    bne _inv_dma_have
+    jmp.w _no_inv_dma
+_inv_dma_have:
+    ldx.w #0xFFFF
+_inv_dma_find:
+    inx
+    lsr
+    bcc _inv_dma_find
+    cpx.w #0x0006  ; safety: out-of-range index -> no DMA
+    bcs _inv_dma_done
+; X = slot index 0..5. Clear its bit before DMA so re-entry is safe.
+    phx
+    lda.l _dma_slot_bit_lut, x
+    eor.b #0xFF
+    and.l battle_render.dma_dirty_slots
+    sta.l battle_render.dma_dirty_slots
+    plx
+; Build X = slot*2 with X.hi clean (we're in M=8 X=16, plain tax leaks A.hi).
+    rep #0x20
+    txa
+    and.w #0x000F
+    asl
+    tax
+    sep #0x20
+; LUT reads use `.l` (24-bit long) since DBR in NMI isn't bank-20.
+    lda.l _dma_slot_vram_lut, x
+    sta.b 0x0c
+    lda.l _dma_slot_vram_lut + 1, x
+    sta.b 0x0d
+    lda.l _dma_slot_src_lut, x
+    sta.b 0x0a
+    lda.l _dma_slot_src_lut + 1, x
+    sta.b 0x0b
+    rep #0x20
+    lda.b 0x0c
+    tay
+    lda.b 0x0a
+    tax
+    lda.w #0x00A0
+    sta.b 0x0e
+    sep #0x20
+    lda.b #0x70
+    jsr.w _sram_dma_transfer_7
+_inv_dma_done:
+_no_inv_dma:
+    plp
+    }
     lda.l battle_render.pending_transfer_mask
     bit #1
     beq _no_transfer
@@ -934,8 +1488,18 @@ normal length, no visible black strip.
     .if SMART {
     ldx.w #0x400
     } else {
+    .if BATTLE_ITEMS_VWF {
+; Inventory region extends past the legacy 0xC00 window. Bump to
+; 0x1000 so the DMA covers the full BG3 CHR upper half ($B000..$BFFF
+; = tile_ids 0x100..0x1FF). Gains 16 tile_ids for the inventory slot
+; budget at zero risk -- the trailing 0x100 bytes were unused.
+; TODO: don't transfer the whole thing every frame ; only dirty
+; regions need flushing.
+    ldx.w #0x1000
+    } else {
     ldx.w #0xc00
     }
+    }
     stx 0x0e
     plx
     lda #0x70
@@ -1066,4 +1630,22 @@ bits_left_on_tile to 8, and advance the tilemap offset by one row (16 tiles).
 ;02/A64A: A8           TAY
 ;02/A64B: E2 20        SEP #$20
 ;02/A64D: 60           RTS
+
+; Per-slot inventory CHR DMA tables. Auto-generated via .for.
+; - bit_lut[N]    = 1 << N (slot N's dirty-mask bit)
+; - vram_lut[N]   = ($BC00 + N*$A0) >> 1 (word-address into BG3 CHR upper half)
+; - src_lut[N]    = $3000 + (0xC0 + N*10)*16 = $3C00 + N*$A0 (bank-70 offset)
+; Slot N owns 10 tile_ids starting at low byte (0xC0 + N*10), so 160-byte slice.
+_dma_slot_bit_lut:
+    .for k := 0, 6 {
+    .db ( 1 << k )
+    }
+_dma_slot_vram_lut:
+    .for k := 0, 6 {
+    .dw ( 0xBC00 + k * 0xA0 ) >> 1
+    }
+_dma_slot_src_lut:
+    .for k := 0, 6 {
+    .dw 0x3C00 + k * 0xA0
+    }
 }
diff --git a/src/battle/monsters_reloc.s b/src/battle/monsters_reloc.s
index 348ae25..eb7c06f 100644
--- a/src/battle/monsters_reloc.s
+++ b/src/battle/monsters_reloc.s
@@ -1,50 +1,55 @@
 """
+.include "../bank20.i"
+
 Relocated battle monster-name pointer resolver: indexes `assets_monsters_long_ptr[A*2]` and renders the
 resulting string into the active slot.
 """
 .extern assets_monsters_long_ptr
 
-load_monster_pointer:
-"""Resolve a long-form monster name pointer from `assets_monsters_long_ptr[A*2]` and render it into the current slot."""
-    rep #0x20
-    ;lda.w #127 - 3
-    asl
-    tax
-    lda.l assets_monsters_long_ptr, x
-    tax
-    tdc
-    sep #0x20
-    phy
-    jsr.w _initialize_monster_slot_near
-    ply
-    rtl
-
-initialize_monster_slot:
-"""Cross-bank (RTL) entry that clears a monster's display slot to spaces."""
-    jsr.w _initialize_monster_slot_near
-    rtl
-
-tab_escape_code:
-"""Text-stream escape that draws a column of spaces to advance the cursor."""
-    jsr.w _draw_spaces
-    rtl
-
-_initialize_monster_slot_near:
-
-; clear monster slot with spaces
-    lda #11
-    sta 0x00
-
-_draw_spaces:
-    lda.b #0xff
-    sta (0x32), y
-    sta (0x34), y
-    iny
-    lda 0x36
-    sta (0x32), y
-    sta (0x34), y
-    iny
-    dec 0x00
-    bne _draw_spaces
-
-    rts
+
+.alloc battle_monsters_reloc_block in bank20_reloc {
+    load_monster_pointer:
+    """Resolve a long-form monster name pointer from `assets_monsters_long_ptr[A*2]` and render it into the current slot."""
+        rep #0x20
+        ;lda.w #127 - 3
+        asl
+        tax
+        lda.l assets_monsters_long_ptr, x
+        tax
+        tdc
+        sep #0x20
+        phy
+        jsr.w _initialize_monster_slot_near
+        ply
+        rtl
+
+    initialize_monster_slot:
+    """Cross-bank (RTL) entry that clears a monster's display slot to spaces."""
+        jsr.w _initialize_monster_slot_near
+        rtl
+
+    tab_escape_code:
+    """Text-stream escape that draws a column of spaces to advance the cursor."""
+        jsr.w _draw_spaces
+        rtl
+
+    _initialize_monster_slot_near:
+
+    ; clear monster slot with spaces
+        lda #11
+        sta 0x00
+
+    _draw_spaces:
+        lda.b #0xff
+        sta (0x32), y
+        sta (0x34), y
+        iny
+        lda 0x36
+        sta (0x32), y
+        sta (0x34), y
+        iny
+        dec 0x00
+        bne _draw_spaces
+
+        rts
+}
diff --git a/src/battle/redraw_gates.s b/src/battle/redraw_gates.s
index de728a5..910b3f1 100644
--- a/src/battle/redraw_gates.s
+++ b/src/battle/redraw_gates.s
@@ -1,4 +1,6 @@
 """
+.include "../bank20.i"
+
 Battle redraw-gate state + writer helpers.
 
 FF6-style dirty-bit gating for the battle redraw chain. One byte at
@@ -23,290 +25,314 @@ battle_menu_dirty := 0x7EEF9A  ; bit 5 = cmd window, bit 6 = status strip,
 battle_monster_dirty := 0x7EEF9B  ; bits 0-7 = per-monster-slot name redraw
 scp_active_slot := 0x7EEF9C  ; scratch for set_active_char_palette
 scp_pal_byte := 0x7EEF9D  ; current palette byte being applied
-status_hash := 0x7EEF9E  ; hash of char-status bytes; gates DrawStatusText
-obj_names_hash := 0x7EEF9F  ; hash of monster slots + $1822; gates DrawObjNames
-char_hp_hash := 0x7EEFA0  ; hash of char HP bytes; gates DrawCharHP
 
-CMD_DIRTY_BIT := 0x20  ; bit 5 of battle_menu_dirty
-NAMES_DIRTY_BIT := 0x10  ; bit 4 of battle_menu_dirty (char names region)
-MONSTER_DIRTY_BIT := 0x01  ; bit 0 of battle_monster_dirty (any monster name)
+.alloc battle_redraw_gates_block in bank20_reloc {
+    status_hash := 0x7EEF9E  ; hash of char-status bytes; gates DrawStatusText
+    obj_names_hash := 0x7EEF9F  ; hash of monster slots + $1822; gates DrawObjNames
+    char_hp_hash := 0x7EEFA0  ; hash of char HP bytes; gates DrawCharHP
 
-mark_cmd_dirty:
-"""
-Set the cmd-window dirty bit. Callable from any bank via JSL/RTL.
-65816 `tsb` has no long-addressing form  ; emulate via lda/ora/sta.
-"""
-    lda.l battle_menu_dirty
-    ora.b #CMD_DIRTY_BIT
-    sta.l battle_menu_dirty
-    rtl
+    CMD_DIRTY_BIT := 0x20  ; bit 5 of battle_menu_dirty
+    NAMES_DIRTY_BIT := 0x10  ; bit 4 of battle_menu_dirty (char names region)
+    MONSTER_DIRTY_BIT := 0x01  ; bit 0 of battle_monster_dirty (any monster name)
 
-walker_rtl:
-"""
-RTL wrapper around `set_active_char_palette` so bank-02 callers
-can JSL into it with matching pop. Reads $1822 into A first so
-caller doesn't have to set it up.
-"""
-    lda.l 0x7E1822
-    jsr.w set_active_char_palette
-    rtl
+    mark_cmd_dirty:
+    """
+    Set the cmd-window dirty bit. Callable from any bank via JSL/RTL.
+    65816 `tsb` has no long-addressing form  ; emulate via lda/ora/sta.
+    """
+        lda.l battle_menu_dirty
+        ora.b #CMD_DIRTY_BIT
+        sta.l battle_menu_dirty
+        rtl
 
-gate_status_check:
-"""
-Bank-20 body for the DrawStatusText hash gate. XOR of char-slot
-status-1 bytes ($2003+slot*$40). Sets carry on dirty, clears on
-clean. Caller (bank-02 trampoline at $02:97F8) tail-jumps to
-$A2A1 on dirty, rts on clean.
-"""
-    lda.l 0x7E2003
-    eor.l 0x7E2043
-    eor.l 0x7E2083
-    eor.l 0x7E20C3
-    eor.l 0x7E2103
-    cmp.l status_hash
-    beq _gsc_clean
-    sta.l status_hash
-    sec
-    rtl
-
-_gsc_clean:
-    clc
-    rtl
-
-gate_obj_names_check:
-"""
-Bank-20 body for the DrawObjNames hash gate. XOR of monster slot
-type bytes ($29B5..$29B8) + active-char index ($1822). Sets
-carry on dirty (re-render needed), clears carry on clean.
-Caller (bank-02 trampoline at $02:97C2) tail-jumps to $99D3 on
-dirty, rts on clean.
-"""
-    lda.l 0x7E29B5
-    eor.l 0x7E29B6
-    eor.l 0x7E29B7
-    eor.l 0x7E29B8
-    eor.l 0x7E1822
-    cmp.l obj_names_hash
-    beq _goc_clean
-    sta.l obj_names_hash
-    sec
-    rtl
-
-_goc_clean:
-    clc
-    rtl
-
-mark_monsters_dirty_and_init:
-"""
-Hook shim for `UpdateDead` entry at $03:B1A0. The original 4 bytes
-(`tdc  ; tax; stx $a9`) get replaced by a JSL here  ; this helper
-sets the monsters-region dirty bit, then replicates the clobbered
-prelude so execution can fall through to the original loop at
-$03:B1A4. UpdateDead is the engine path that marks a monster
-dead (`sta $29b5,x` with $FF after status apply)  ; flagging
-monsters dirty here lets the gated monster-name trampoline
-re-render once the dead slot is wiped.
-"""
-    lda.l battle_render.region_dirty_bits
-    ora.b #battle_render.REGION_DIRTY_MONSTERS
-    sta.l battle_render.region_dirty_bits
-    tdc
-    tax
-    stx.b 0xa9
-    rtl
-
-mark_all_dirty:
-"""
-Reset both dirty bytes to $FF so the next frame renders everything.
-Called once at battle init.
-"""
-    lda.b #0xFF
-    sta.l battle_menu_dirty
-    sta.l battle_monster_dirty
-    rtl
-
-    .extern clear_names_window_buffer
-    .extern battle_render.REGION_DIRTY_MONSTERS
-    .extern battle_render.REGION_DIRTY_NAMES
-    .extern battle_render.region_dirty_bits
-    .extern battle_render.render_skipped
-    .extern battle_render.tilemap_pending_mask
-    .extern battle_render.TILEMAP_PENDING_MAIN
-
-reset_queue_dirty_bits:
-"""
-Seed all redraw-gate state for a fresh battle. Called once per
-battle from the InitMenuWindows hook ($02:9A63). Normal sense
-everywhere: 1 = dirty, 0 = clean. Seed all bytes to $FF so the
-first frame renders everything  ; the per-region gates clear their
-own bits after rendering, and writer sites re-arm on state change.
-"""
-    lda.b #0x00
-    sta.l battle_render.render_skipped
-    lda.b #0xFF
-    sta.l battle_render.region_dirty_bits
-    sta.l battle_menu_dirty
-    sta.l battle_monster_dirty
-    rtl
-
-gated_clear_names_window_buffer:
-"""
-Wrap `clear_names_window_buffer` ($02:A299 call site) with the
-names-region dirty-bit check. When names is clean (bit clear),
-skip the tilemap wipe so the gated `init_names_gated` + skipped
-DrawText leaves the VRAM tilemap untouched. Without this the
-tilemap wipe still fires every frame and the gated render skip
-leaves blank tiles on screen.
-"""
-    lda.l battle_render.region_dirty_bits
-    bit.b #battle_render.REGION_DIRTY_NAMES
-    beq _gcnwb_skip
-    jsr.l clear_names_window_buffer
+    walker_rtl:
+    """
+    RTL wrapper around `set_active_char_palette` so bank-02 callers
+    can JSL into it with matching pop. Reads $1822 into A first so
+    caller doesn't have to set it up.
+    """
+        lda.l 0x7E1822
+        jsr.w set_active_char_palette
+        rtl
 
-_gcnwb_skip:
-    rtl
+    gate_status_check:
+    """
+    Bank-20 body for the DrawStatusText hash gate. XOR of char-slot
+    status-1 bytes ($2003+slot*$40). Sets carry on dirty, clears on
+    clean. Caller (bank-02 trampoline at $02:97F8) tail-jumps to
+    $A2A1 on dirty, rts on clean.
+    """
+        lda.l 0x7E2003
+        eor.l 0x7E2043
+        eor.l 0x7E2083
+        eor.l 0x7E20C3
+        eor.l 0x7E2103
+        cmp.l status_hash
+        beq _gsc_clean
+        sta.l status_hash
+        sec
+        rtl
 
-set_active_char_palette:
-"""
-Walk all 5 char-name slots in the `$7E:B966` tilemap. Active slot
-gets palette $04 (bit 2 of tilemap-entry hi byte = palette 1)  ;
-others get $00 (palette 0).
-
-Each char-name slot occupies row at `$B966 + slot * $40`. Names
-are max 6 tiles wide in the FR translation. Each tilemap entry
-is 2 bytes (low = tile id, hi = flags+palette). Two pointer
-mirrors get written by the VWF dispatcher (`tilemap_write`
-in message.s) at offsets +0 and +$0C of the row base, so 12 hi
-bytes total per slot (6 on each mirror).
-
-A = active slot index (0..4) on entry. M=8, X=8.
-"""
+    _gsc_clean:
+        clc
+        rtl
 
+    gate_obj_names_check:
+    """
+    Bank-20 body for the DrawObjNames hash gate. XOR of monster slot
+    type bytes ($29B5..$29B8) + active-char index ($1822). Sets
+    carry on dirty (re-render needed), clears carry on clean.
+    Caller (bank-02 trampoline at $02:97C2) tail-jumps to $99D3 on
+    dirty, rts on clean.
+    """
+        lda.l 0x7E29B5
+        eor.l 0x7E29B6
+        eor.l 0x7E29B7
+        eor.l 0x7E29B8
+        eor.l 0x7E1822
+        cmp.l obj_names_hash
+        beq _goc_clean
+        sta.l obj_names_hash
+        sec
+        rtl
 
-    php
-    sep #0x20
-    rep #0x10
-    phb
-    pha
-    lda.b #0x7E  ; force DBR = $7E so `(0x32),y` writes to WRAM
-    pha
-    plb
-    pla
-    sta.l scp_active_slot
-    ; Save $32/$33 ; walker reuses as scratch indirect-ptr ; NMI
-    ; caller's BG / DMA state needs them preserved.
-    rep #0x20
-    lda.b 0x32
-    pha
-    sep #0x20
-    ldx.w #0
-
-_scp_slot_loop:
-    cpx.w #5
-    bcs _scp_done
-    txa
-    cmp.l scp_active_slot
-    beq _scp_is_active
-    lda #0x00
-    bra _scp_have_pal
-
-_scp_is_active:
-    lda #0x08  ; palette 2 (vanilla `lda #$08` in UpdateCharNames @a24c)
-
-_scp_have_pal:
-    sta.l scp_pal_byte
-
-; base = $B966 + slot * 24 (per-slot stride confirmed via trace:
-; $32 mirror at slot*24+0 (6 tiles padding), $34 mirror at
-; slot*24+12 (6 tiles real name content)).
-    rep #0x20
-    txa
-    and.w #0x000F
-    asl
-    asl
-    asl  ; *8
-    pha
-    asl  ; *16
-    clc
-    adc 1, s  ; *16 + *8 = *24
-    clc
-    adc.w #0xB966
-    sta.b 0x32
-    pla  ; balance stack
-    sep #0x20
-    ; Patch 6 entries on the ($32) mirror at offsets +1, +3, +5, +7, +9, +B
-    ldy.w #1
-    jsr.w _scp_patch_six
-    ; Now patch the ($34) mirror at base + $0C
-    rep #0x20
-    lda.b 0x32
-    clc
-    adc.w #0x000C
-    sta.b 0x32
-    sep #0x20
-    ldy.w #1
-    jsr.w _scp_patch_six
-    inx
-    bra _scp_slot_loop
-
-_scp_done:
-    rep #0x20
-    pla
-    sta.b 0x32
-    sep #0x20
-    plb
-    plp
-    rts
-
-_scp_patch_six:
-    ; ($32) = row base in bank $7E. Y = first hi-byte offset.
-    ; Walks Y, Y+2, ..., Y+10. Masks `$E3`, ORs in `scp_pal_byte`.
-    phx
-    ldx.w #6
-
-_scp_loop_six:
-    lda (0x32), y
-    and #0xE3
-    ora.l scp_pal_byte
-    sta (0x32), y
-    iny
-    iny
-    dex
-    bne _scp_loop_six
-    plx
-    rts
-
-set_active_char_and_dirty:
-"""
-Writer-site shim for the active-char store (`sta $1822` at $03:A482,
-followed by `sta $d0`). Original 5 bytes replaced by JSL + NOP. Shim
-performs both stores then unconditionally marks cmd + names +
-monsters dirty. The vanilla site fires only when the battle-menu
-queue pops a new char (`check if menu needs to open` path in
-`battle/char.asm:464`), so re-arming on every call is correct  ;
-a compare-against-current would mis-skip when the popped char
-matches a stale value in $1822, leaving the prior cmd-window
-tilemap on screen.
-Caller has M=8, A holds the char index, DBR may differ from $7E.
-"""
-    sep #0x20
-    sta.l 0x7E1822
-    sta.l 0x7E00D0
-    pha
-    lda.l battle_menu_dirty
-    ora.b #CMD_DIRTY_BIT
-    sta.l battle_menu_dirty
-    ; Char-name palette patch (replaces the full names VWF re-render
-    ; that used to fire via REGION_DIRTY_NAMES on every rotation).
-    ; Walks the `$7E:B966` tilemap, rewrites the palette field of
-    ; each char's 6 max tiles on both pointer mirrors ; ~300 cycles
-    ; vs ~3M for the VWF re-render. Also flip names tilemap dirty
-    ; so the unified-DMA path uploads the patched tilemap to VRAM.
-    pla
-    pha
-    jsr.w set_active_char_palette
-    lda.l battle_render.tilemap_pending_mask
-    ora.b #battle_render.TILEMAP_PENDING_MAIN
-    sta.l battle_render.tilemap_pending_mask
-    pla
-    rtl
+    _goc_clean:
+        clc
+        rtl
+
+    mark_monsters_dirty_and_init:
+    """
+    Hook shim for `UpdateDead` entry at $03:B1A0. The original 4 bytes
+    (`tdc  ; tax; stx $a9`) get replaced by a JSL here  ; this helper
+    sets the monsters-region dirty bit, then replicates the clobbered
+    prelude so execution can fall through to the original loop at
+    $03:B1A4. UpdateDead is the engine path that marks a monster
+    dead (`sta $29b5,x` with $FF after status apply)  ; flagging
+    monsters dirty here lets the gated monster-name trampoline
+    re-render once the dead slot is wiped.
+    """
+        lda.l battle_render.region_dirty_bits
+        ora.b #battle_render.REGION_DIRTY_MONSTERS
+        sta.l battle_render.region_dirty_bits
+        ; Propagate to cmd-window region: the cmd-window tilemap at $C1A5+ is
+        ; a mirror of the main view ($BE65+) overlaid with cmd tiles. If we
+        ; refresh monsters in the main view, the cmd mirror is stale, so set
+        ; CMD_DIRTY_BIT here too. The relocated DrawCmdWindow path picks this
+        ; up next frame and re-runs the WRAM mirror via `mirror_main_to_cmd`.
+        lda.l battle_menu_dirty
+        ora.b #CMD_DIRTY_BIT
+        sta.l battle_menu_dirty
+        tdc
+        tax
+        stx.b 0xa9
+        rtl
+
+    mark_all_dirty:
+    """
+    Reset both dirty bytes to $FF so the next frame renders everything.
+    Called once at battle init.
+    """
+        lda.b #0xFF
+        sta.l battle_menu_dirty
+        sta.l battle_monster_dirty
+        rtl
+
+        .extern clear_names_window_buffer
+        .extern battle_render.REGION_DIRTY_MONSTERS
+        .extern battle_render.REGION_DIRTY_NAMES
+        .extern battle_render.region_dirty_bits
+        .extern battle_render.render_skipped
+        .extern battle_render.tilemap_pending_mask
+        .extern battle_render.TILEMAP_PENDING_MAIN
+
+    reset_queue_dirty_bits:
+    """
+    Seed all redraw-gate state for a fresh battle. Called once per
+    battle from the InitMenuWindows hook ($02:9A63). Normal sense
+    everywhere: 1 = dirty, 0 = clean. Seed all bytes to $FF so the
+    first frame renders everything  ; the per-region gates clear their
+    own bits after rendering, and writer sites re-arm on state change.
+    """
+        lda.b #0x00
+        sta.l battle_render.render_skipped
+        lda.b #0xFF
+        sta.l battle_render.region_dirty_bits
+        sta.l battle_menu_dirty
+        sta.l battle_monster_dirty
+        rtl
+
+    gated_clear_names_window_buffer:
+    """
+    Wrap `clear_names_window_buffer` ($02:A299 call site) with the
+    names-region dirty-bit check. When names is clean (bit clear),
+    skip the tilemap wipe so the gated `init_names_gated` + skipped
+    DrawText leaves the VRAM tilemap untouched. Without this the
+    tilemap wipe still fires every frame and the gated render skip
+    leaves blank tiles on screen.
+    """
+        lda.l battle_render.region_dirty_bits
+        bit.b #battle_render.REGION_DIRTY_NAMES
+        beq _gcnwb_skip
+        jsr.l clear_names_window_buffer
+
+    _gcnwb_skip:
+        rtl
+
+    set_active_char_palette:
+    """
+    Walk all 5 char-name slots in the `$7E:B966` tilemap. Active slot
+    gets palette $04 (bit 2 of tilemap-entry hi byte = palette 1)  ;
+    others get $00 (palette 0).
+
+    Each char-name slot occupies row at `$B966 + slot * $40`. Names
+    are max 6 tiles wide in the FR translation. Each tilemap entry
+    is 2 bytes (low = tile id, hi = flags+palette). Two pointer
+    mirrors get written by the VWF dispatcher (`tilemap_write`
+    in message.s) at offsets +0 and +$0C of the row base, so 12 hi
+    bytes total per slot (6 on each mirror).
+
+    A = active character SLOT on entry (vanilla `$1822` semantics).
+
+    The on-screen tilemap is ordered by display position, NOT slot:
+    row R of `$B966` shows whoever vanilla's `CharOrderTbl`
+    (`$02:A1C8` = .byte 1,3,0,4,2) maps R to. The per-row loop below
+    already iterates display rows 0..4, so we compare each row's
+    CharOrderTbl entry against $1822 instead of comparing the row
+    index itself. Without this indirection, Cecil acting (slot 0)
+    used to light up row 0 of the tilemap, which displays slot 1
+    (= Palom in the default order).
+
+    M=8, X=8.
+    """
+
+
+        php
+        sep #0x20
+        rep #0x10
+        phb
+        pha
+        lda.b #0x7E  ; force DBR = $7E so `(0x32),y` writes to WRAM
+        pha
+        plb
+        pla
+        sta.l scp_active_slot
+        ; Save $32/$33 ; walker reuses as scratch indirect-ptr ; NMI
+        ; caller's BG / DMA state needs them preserved.
+        rep #0x20
+        lda.b 0x32
+        pha
+        sep #0x20
+        ldx.w #0
+
+    _scp_slot_loop:
+        cpx.w #5
+        bcs _scp_done
+        ; CharOrderTbl[row] = char slot displayed at this row. Compare to
+        ; the active slot; match -> highlight palette, miss -> palette 0.
+        lda.l 0x02A1C8, x
+        cmp.l scp_active_slot
+        beq _scp_is_active
+        lda #0x00
+        bra _scp_have_pal
+
+    _scp_is_active:
+        lda #0x08  ; palette 2 (vanilla `lda #$08` in UpdateCharNames @a24c)
+
+    _scp_have_pal:
+        sta.l scp_pal_byte
+
+    ; base = $B966 + slot * 24 (per-slot stride confirmed via trace:
+    ; $32 mirror at slot*24+0 (6 tiles padding), $34 mirror at
+    ; slot*24+12 (6 tiles real name content)).
+        rep #0x20
+        txa
+        and.w #0x000F
+        asl
+        asl
+        asl  ; *8
+        pha
+        asl  ; *16
+        clc
+        adc 1, s  ; *16 + *8 = *24
+        clc
+        adc.w #0xB966
+        sta.b 0x32
+        pla  ; balance stack
+        sep #0x20
+        ; Patch 6 entries on the ($32) mirror at offsets +1, +3, +5, +7, +9, +B
+        ldy.w #1
+        jsr.w _scp_patch_six
+        ; Now patch the ($34) mirror at base + $0C
+        rep #0x20
+        lda.b 0x32
+        clc
+        adc.w #0x000C
+        sta.b 0x32
+        sep #0x20
+        ldy.w #1
+        jsr.w _scp_patch_six
+        inx
+        bra _scp_slot_loop
+
+    _scp_done:
+        rep #0x20
+        pla
+        sta.b 0x32
+        sep #0x20
+        plb
+        plp
+        rts
+
+    _scp_patch_six:
+        ; ($32) = row base in bank $7E. Y = first hi-byte offset.
+        ; Walks Y, Y+2, ..., Y+10. Masks `$E3`, ORs in `scp_pal_byte`.
+        phx
+        ldx.w #6
+
+    _scp_loop_six:
+        lda (0x32), y
+        and #0xE3
+        ora.l scp_pal_byte
+        sta (0x32), y
+        iny
+        iny
+        dex
+        bne _scp_loop_six
+        plx
+        rts
+
+    set_active_char_and_dirty:
+    """
+    Writer-site shim for the active-char store (`sta $1822` at $03:A482,
+    followed by `sta $d0`). Original 5 bytes replaced by JSL + NOP. Shim
+    performs both stores then unconditionally marks cmd + names +
+    monsters dirty. The vanilla site fires only when the battle-menu
+    queue pops a new char (`check if menu needs to open` path in
+    `battle/char.asm:464`), so re-arming on every call is correct  ;
+    a compare-against-current would mis-skip when the popped char
+    matches a stale value in $1822, leaving the prior cmd-window
+    tilemap on screen.
+    Caller has M=8, A holds the char index, DBR may differ from $7E.
+    """
+        sep #0x20
+        sta.l 0x7E1822
+        sta.l 0x7E00D0
+        pha
+        lda.l battle_menu_dirty
+        ora.b #CMD_DIRTY_BIT
+        sta.l battle_menu_dirty
+        ; Char-name palette patch (replaces the full names VWF re-render
+        ; that used to fire via REGION_DIRTY_NAMES on every rotation).
+        ; Walks the `$7E:B966` tilemap, rewrites the palette field of
+        ; each char's 6 max tiles on both pointer mirrors ; ~300 cycles
+        ; vs ~3M for the VWF re-render. Also flip names tilemap dirty
+        ; so the unified-DMA path uploads the patched tilemap to VRAM.
+        pla
+        pha
+        jsr.w set_active_char_palette
+        lda.l battle_render.tilemap_pending_mask
+        ora.b #battle_render.TILEMAP_PENDING_MAIN
+        sta.l battle_render.tilemap_pending_mask
+        pla
+        rtl
+}
diff --git a/src/battle/redraw_writer_patches.s b/src/battle/redraw_writer_patches.s
index a06e1da..ceba5a1 100644
--- a/src/battle/redraw_writer_patches.s
+++ b/src/battle/redraw_writer_patches.s
@@ -92,112 +92,125 @@ follow-up patches.
 
 .pool bank02_battle_redraw_helpers {
     range 0x0297AB 0x029824
-    fill 0
     strategy order
 }
 
 .alloc battle_redraw_helpers in bank02_battle_redraw_helpers {
-    ; --- Names + monster gated trampolines (slice 2, queue-side bits) ---
-    ; The init -> DrawText -> deinit pipeline still fires every frame so
-    ; battle_flags symmetry is preserved (other VWF callers like HP/MP
-    ; digits keep seeing a consistent dispatcher state). The gating is
-    ; INSIDE `init_*_gated`: it reads the region's clean bit from
-    ; `region_dirty_bits` at $703C01. If clean, it sets `render_skipped`
-    ; ($703C02) to $FF; the trampoline reads that and skips DrawText, and
-    ; `deinit_gated` skips the DMA-signal. The WRAM tile buffer is left
-    ; untouched, no DMA fires from this region's deinit, and the VRAM
-    ; tilemap persists from the previous render.
-    _msg_monster_window_gated:
-    """Gated DrawMonsterNames trampoline (slice-2 queue-side bit)."""
+; --- Names + monster gated trampolines (slice 2, queue-side bits) ---
+; The init -> DrawText -> deinit pipeline still fires every frame so
+; battle_flags symmetry is preserved (other VWF callers like HP/MP
+; digits keep seeing a consistent dispatcher state). The gating is
+; INSIDE `init_*_gated`: it reads the region's clean bit from
+; `region_dirty_bits` at $703C01. If clean, it sets `render_skipped`
+; ($703C02) to $FF; the trampoline reads that and skips DrawText, and
+; `deinit_gated` skips the DMA-signal. The WRAM tile buffer is left
+; untouched, no DMA fires from this region's deinit, and the VRAM
+; tilemap persists from the previous render.
+_msg_monster_window_gated:
+"""Gated DrawMonsterNames trampoline (slice-2 queue-side bit)."""
     jsr.l messages_vwf.init_monsters_gated
     lda.l battle_render.render_skipped
     bne _mmwg_after_draw
     jsr 0xA455
-    ; DrawText
-    _mmwg_after_draw:
+; DrawText
+_mmwg_after_draw:
     jsr.l messages_vwf.deinit_gated
     rts
-    ; --- Bank-02 trampoline for NMI-side UpdateFlyingHDMA ---
-    ; `dma_transfer` (bank-20) calls this via JSL ; trampoline JSR's
-    ; into UpdateFlyingHDMA ($02:82E1, phase-2 spin-wait nopped) and
-    ; RTLs back. Matches the JSL push/pop convention ; the vanilla
-    ; function ends in rts, our trampoline rtl's instead.
-    flying_hdma_trampoline:
-    """Bank-02 RTL wrapper around vanilla `UpdateFlyingHDMA` ($02:82E1)."""
+; --- Bank-02 trampoline for NMI-side UpdateFlyingHDMA ---
+; `dma_transfer` (bank-20) calls this via JSL ; trampoline JSR's
+; into UpdateFlyingHDMA ($02:82E1, phase-2 spin-wait nopped) and
+; RTLs back. Matches the JSL push/pop convention ; the vanilla
+; function ends in rts, our trampoline rtl's instead.
+flying_hdma_trampoline:
+"""
+Bank-02 RTL wrapper around vanilla `UpdateFlyingHDMA` ($02:82E1).
+    NMI caller leaves M=X=16  ; vanilla expects M=X=8 (main-loop
+    convention). Without forcing M=8 the routine reads/writes 16-bit
+    through `$ED4E`, `$1813`, and the buffer fill at `$7614,x`, so
+    every BG vscroll entry gets 0x0000 and the flying animation
+    freezes the moment our NMI dispatch takes over.
+"""
+
+
+    php
+    rep #0x30
+    lda.w #0x0000
+    sep #0x20
     jsr 0x82E1
+    plp
     rtl
-    ; --- Battle-init seed: zero/seed all redraw-gate state at the
-    ; `Battle_ext` root entry ($03:8000). Runs exactly once per battle
-    ; (from field.asm:1005 `jsl Battle_ext` and menu.asm:163
-    ; `jml Battle_ext`  ; no other callers per ff4decomp). Earlier we
-    ; tried `InitMenuWindows` ($02:9A63) but that ran INSIDE the
-    ; UpdateCharNames per-char loop entry, and a 2nd-battle kss restore
-    ; mid-state could skip it entirely. The root entry is the only
-    ; correct seed point.
-    ;
-    ; Original `Battle_ext` was `jmp ExecBattle` (3 bytes). JML is 4
-    ; bytes so we clobber 1 byte of `Battle_ext2` at $03:8003. Decomp
-    ; grep confirms zero callers of Battle_ext2 anywhere.
-    _battle_ext_seed:
-    """Battle_ext root tail: seed redraw-gate state then jump ExecBattle."""
+; --- Battle-init seed: zero/seed all redraw-gate state at the
+; `Battle_ext` root entry ($03:8000). Runs exactly once per battle
+; (from field.asm:1005 `jsl Battle_ext` and menu.asm:163
+; `jml Battle_ext`  ; no other callers per ff4decomp). Earlier we
+; tried `InitMenuWindows` ($02:9A63) but that ran INSIDE the
+; UpdateCharNames per-char loop entry, and a 2nd-battle kss restore
+; mid-state could skip it entirely. The root entry is the only
+; correct seed point.
+;
+; Original `Battle_ext` was `jmp ExecBattle` (3 bytes). JML is 4
+; bytes so we clobber 1 byte of `Battle_ext2` at $03:8003. Decomp
+; grep confirms zero callers of Battle_ext2 anywhere.
+_battle_ext_seed:
+"""Battle_ext root tail: seed redraw-gate state then jump ExecBattle."""
     jsr.l reset_queue_dirty_bits
     jmp.l 0x038009
-    ; ExecBattle
-    ; --- DrawStatusText gate (hash-based) ---
-    ; RedrawMainMenu @96C8 = `jsr DrawStatusText` ; 9.33M cycles per 60f
-    ; (top remaining hitter after the cmd-window gate). Skip when char
-    ; status state is unchanged. Hash = XOR of `$2003+slot*$40` for the
-    ; 5 char slots (status 1 byte). Cached at `status_hash` ; first call
-    ; per battle always renders (cache initialized to 0 by Battle_ext
-    ; seed but state hash != 0 in normal play). Status flicker pulse
-    ; freezes when no status changes ; acceptable trade for ~9M cycles.
-    gate_draw_status_text:
-    """Bank-02 trampoline  ; JSL gate_status_check, jmp $A2A1 on dirty, rts on clean."""
+; ExecBattle
+; --- DrawStatusText gate (hash-based) ---
+; RedrawMainMenu @96C8 = `jsr DrawStatusText` ; 9.33M cycles per 60f
+; (top remaining hitter after the cmd-window gate). Skip when char
+; status state is unchanged. Hash = XOR of `$2003+slot*$40` for the
+; 5 char slots (status 1 byte). Cached at `status_hash` ; first call
+; per battle always renders (cache initialized to 0 by Battle_ext
+; seed but state hash != 0 in normal play). Status flicker pulse
+; freezes when no status changes ; acceptable trade for ~9M cycles.
+gate_draw_status_text:
+"""Bank-02 trampoline  ; JSL gate_status_check, jmp $A2A1 on dirty, rts on clean."""
     jsr.l gate_status_check
     bcc _gdst_skip
     jmp 0xA2A1
-    _gdst_skip:
+_gdst_skip:
     rts
-    ; --- DrawObjNames gate (hash of monster slots + $1822) ---
-    ; Bank-20 body returns carry-set when re-render needed ; bank-02
-    ; trampoline tail-jumps to vanilla DrawObjNames on dirty, rts on
-    ; clean. Earlier we noop'd UpdateMagicList ($02:96D4) to reclaim
-    ; 36 bytes for this gate's body, but that broke Rydia's summon
-    ; menu display (vanilla UpdateMagicList still needed for periodic
-    ; magic list refresh ; ImmediateMenuUpdate alone doesn't cover
-    ; all dispatch paths). Restored.
-    gate_draw_obj_names:
-    """Bank-02 trampoline  ; JSL gate_obj_names_check, jmp $99D3 on dirty, rts on clean."""
+; --- DrawObjNames gate (hash of monster slots + $1822) ---
+; Bank-20 body returns carry-set when re-render needed ; bank-02
+; trampoline tail-jumps to vanilla DrawObjNames on dirty, rts on
+; clean. Earlier we noop'd UpdateMagicList ($02:96D4) to reclaim
+; 36 bytes for this gate's body, but that broke Rydia's summon
+; menu display (vanilla UpdateMagicList still needed for periodic
+; magic list refresh ; ImmediateMenuUpdate alone doesn't cover
+; all dispatch paths). Restored.
+gate_draw_obj_names:
+"""Bank-02 trampoline  ; JSL gate_obj_names_check, jmp $99D3 on dirty, rts on clean."""
     jsr.l gate_obj_names_check
     bcc _gdon_skip
     jmp 0x99D3
-    _gdon_skip:
+_gdon_skip:
     rts
-    ; --- Battle-init highlight stamp ---
-    ; Reclaim the 3-nop slot at $02:9A69 (previously vanilla `jsr
-    ; InitMagicListTextBuf` ; we noop'd that in magic/patches.s since
-    ; our two-column magic display drives its own buffer). Insert
-    ; `jsr walker_helper` so the active-char palette gets stamped
-    ; right after UpdateCharNames built $B966 ; subsequent init steps
-    ; (DrawHPText, DrawMonsterNames, etc.) target different WRAM
-    ; regions ($B9DE, $BB1E) so the stamp survives.
-    walker_helper:
-    """5-byte bank-02 trampoline: JSL bank-20 walker, RTS to caller."""
+; --- Battle-init highlight stamp ---
+; Reclaim the 3-nop slot at $02:9A69 (previously vanilla `jsr
+; InitMagicListTextBuf` ; we noop'd that in magic/patches.s since
+; our two-column magic display drives its own buffer). Insert
+; `jsr walker_helper` so the active-char palette gets stamped
+; right after UpdateCharNames built $B966 ; subsequent init steps
+; (DrawHPText, DrawMonsterNames, etc.) target different WRAM
+; regions ($B9DE, $BB1E) so the stamp survives.
+walker_helper:
+"""5-byte bank-02 trampoline: JSL bank-20 walker, RTS to caller."""
     jsr.l walker_rtl
     rts
-    _msg_names_window_gated:
-    """Gated DrawCharNames trampoline (slice-2 queue-side bit)."""
+_msg_names_window_gated:
+"""Gated DrawCharNames trampoline (slice-2 queue-side bit)."""
     lda.b 0x4A
     and.b 0x04
     bne _mnwg_done
-    ; inventory open -> skip whole pipeline
+; inventory open -> skip whole pipeline
     jsr.l messages_vwf.init_names_gated
     lda.l battle_render.render_skipped
     bne _mnwg_after_draw
     jsr 0xA455
-    _mnwg_after_draw:
+_mnwg_after_draw:
     jsr.l messages_vwf.deinit_gated
-    _mnwg_done:
+_mnwg_done:
     rts
 }
 ; end .alloc battle_redraw_helpers
diff --git a/src/battle/sram.s b/src/battle/sram.s
index 6512167..e7bba76 100644
--- a/src/battle/sram.s
+++ b/src/battle/sram.s
@@ -4,6 +4,7 @@ put_char/put_char_with_dakuten, `battle_display_char` jump-table dispatch and th
 helper.
 """
 .include "src/battle/sram.i"
+.include "../bank20.i"
 
 .import "assets"
 .import "dakuten"
@@ -11,137 +12,140 @@ helper.
 
 BATTLE_DAKUTEN_TABLE = 0x16FA40
 
-.scope battle_flags {
-    """Battle-flags toggles for switching the message renderer between WRAM tiles and VWF."""
-set_vwf_render:
-"""NOTE: set_sram_copy and clear_sram_copy removed - SRAM mode no longer used"""
-    battle_flags_set(0x02)
-    rtl
-clear_vwf_render:
-    battle_flags_clear(0x02)
-    rtl
-}
 
-copy_battle_char:
-"""Copy a glyph + its dakuten companion from the SRAM staging area to the destination pair."""
-    lda.l sram_base + 0x2E00, x
-    sta (0x00), y
-    lda.l sram_base + 0x2E00 + 0x30, x
-    sta (0x02), y
-    rtl
-
-.scope wram {
-    """WRAM-mode put-char primitives used by the original battle text renderer."""
-put_char:
-    phx
-    sta (0x34), y
-    lda #0xFF
-    sta (0x32), y
-    iny
-    lda 0x36
-    sta (0x32), y
-    sta (0x34), y
-    iny
-    plx
-    rtl
-put_char_with_dakuten:
-    phx
-    .if 0 {
-    sec
-    sbc #0xF
-    asl
-    tax
-    lda.l BATTLE_DAKUTEN_TABLE, x
-    sta (0x32), y
-    lda.l BATTLE_DAKUTEN_TABLE + 1, x
-    sta (0x34), y
-    } else {
-    jsr.l lookup_dakuten
-    sta (0x32), y
-    xba
-    sta (0x34), y
-    lda #0x00
-    xba
+.alloc battle_sram_block in bank20_reloc {
+    .scope battle_flags {
+        """Battle-flags toggles for switching the message renderer between WRAM tiles and VWF."""
+    set_vwf_render:
+    """NOTE: set_sram_copy and clear_sram_copy removed - SRAM mode no longer used"""
+        battle_flags_set(0x02)
+        rtl
+    clear_vwf_render:
+        battle_flags_clear(0x02)
+        rtl
     }
-    iny
-    lda 0x36
-    sta (0x32), y
-    sta (0x34), y
-    iny
-    plx
-    rtl
-}
 
-; NOTE: sram scope removed - SRAM mode no longer used
-; .scope sram { put_char, put_char_with_dakuten }
-
-battle_display_char:
-"""
-Dispatch a fixed-mode char draw to either the WRAM put_char or the messages_vwf renderer based on the active
-battle_flags.
-"""
-{
-    battle_flag_switch(battle_flags_jump_table)
-battle_flags_jump_table:
-    .dw wram.put_char  ; index 0 (flags = 0)
-    .dw wram.put_char  ; index 2 (flags = 1) - fallback to WRAM
-    .dw messages_vwf.put_fixed_char_dakuten_far  ; index 4 (flags = 2)
-    .dw messages_vwf.put_fixed_char_dakuten_far  ; index 6 (flags = 3) - fallback
-}
+    copy_battle_char:
+    """Copy a glyph + its dakuten companion from the SRAM staging area to the destination pair."""
+        lda.l sram_base + 0x2E00, x
+        sta (0x00), y
+        lda.l sram_base + 0x2E00 + 0x30, x
+        sta (0x02), y
+        rtl
+
+    .scope wram {
+        """WRAM-mode put-char primitives used by the original battle text renderer."""
+    put_char:
+        phx
+        sta (0x34), y
+        lda #0xFF
+        sta (0x32), y
+        iny
+        lda 0x36
+        sta (0x32), y
+        sta (0x34), y
+        iny
+        plx
+        rtl
+    put_char_with_dakuten:
+        phx
+        .if 0 {
+        sec
+        sbc #0xF
+        asl
+        tax
+        lda.l BATTLE_DAKUTEN_TABLE, x
+        sta (0x32), y
+        lda.l BATTLE_DAKUTEN_TABLE + 1, x
+        sta (0x34), y
+        } else {
+        jsr.l lookup_dakuten
+        sta (0x32), y
+        xba
+        sta (0x34), y
+        lda #0x00
+        xba
+        }
+        iny
+        lda 0x36
+        sta (0x32), y
+        sta (0x34), y
+        iny
+        plx
+        rtl
+    }
 
-battle_display_dakuten_char:
-"""
-Dakuten-aware variant of `battle_display_char`: routes to the dakuten put_char or to messages_vwf depending on
-battle_flags.
-"""
-{
-    battle_flag_switch(battle_flags_jump_table)
-battle_flags_jump_table:
-    .dw wram.put_char_with_dakuten  ; index 0 (flags = 0)
-    .dw wram.put_char_with_dakuten  ; index 2 (flags = 1) - fallback to WRAM
-    .dw messages_vwf.put_fixed_char_no_dakuten_far  ; index 4 (flags = 2)
-    .dw messages_vwf.put_fixed_char_no_dakuten_far  ; index 6 (flags = 3) - fallback
-}
+    ; NOTE: sram scope removed - SRAM mode no longer used
+    ; .scope sram { put_char, put_char_with_dakuten }
+
+    battle_display_char:
+    """
+    Dispatch a fixed-mode char draw to either the WRAM put_char or the messages_vwf renderer based on the active
+    battle_flags.
+    """
+    {
+        battle_flag_switch(battle_flags_jump_table)
+    battle_flags_jump_table:
+        .dw wram.put_char  ; index 0 (flags = 0)
+        .dw wram.put_char  ; index 2 (flags = 1) - fallback to WRAM
+        .dw messages_vwf.put_fixed_char_dakuten_far  ; index 4 (flags = 2)
+        .dw messages_vwf.put_fixed_char_dakuten_far  ; index 6 (flags = 3) - fallback
+    }
 
-_sink:
-    rtl
+    battle_display_dakuten_char:
+    """
+    Dakuten-aware variant of `battle_display_char`: routes to the dakuten put_char or to messages_vwf depending on
+    battle_flags.
+    """
+    {
+        battle_flag_switch(battle_flags_jump_table)
+    battle_flags_jump_table:
+        .dw wram.put_char_with_dakuten  ; index 0 (flags = 0)
+        .dw wram.put_char_with_dakuten  ; index 2 (flags = 1) - fallback to WRAM
+        .dw messages_vwf.put_fixed_char_no_dakuten_far  ; index 4 (flags = 2)
+        .dw messages_vwf.put_fixed_char_no_dakuten_far  ; index 6 (flags = 3) - fallback
+    }
 
-clear_names_window_buffer:
-"""
-Fill the names-window WRAM tilemap buffer with $FF (transparent / blank tile) starting at the address held in
-$EF52.
-"""
-    phx
-    phy
-    rep #0x20
-    ldy.w #0
-    ldx.w 0xef52
-
-_clear_name_loop:
-    lda.w #0x00ff
-    sta.l 0x7e0000, x
-    sta.l 0x7e0002, x
-    sta.l 0x7e0004, x
-    sta.l 0x7e0006, x
-    sta.l 0x7e0008, x
-    sta.l 0x7e000a, x
-
-    txa
-    clc
-    adc.w #6 * 2
-    tax
-    iny
-    tya
-    cmp.w #5 * 2
-
-    bne _clear_name_loop
-
-
-    sep #0x20
-    ply
-    plx
-    tdc
-    sta 0x74FC, y
-    rtl
-
-    .include "src/battle/message.s"
+    _sink:
+        rtl
+
+    clear_names_window_buffer:
+    """
+    Fill the names-window WRAM tilemap buffer with $FF (transparent / blank tile) starting at the address held in
+    $EF52.
+    """
+        phx
+        phy
+        rep #0x20
+        ldy.w #0
+        ldx.w 0xef52
+
+    _clear_name_loop:
+        lda.w #0x00ff
+        sta.l 0x7e0000, x
+        sta.l 0x7e0002, x
+        sta.l 0x7e0004, x
+        sta.l 0x7e0006, x
+        sta.l 0x7e0008, x
+        sta.l 0x7e000a, x
+
+        txa
+        clc
+        adc.w #6 * 2
+        tax
+        iny
+        tya
+        cmp.w #5 * 2
+
+        bne _clear_name_loop
+
+
+        sep #0x20
+        ply
+        plx
+        tdc
+        sta 0x74FC, y
+        rtl
+
+        .include "src/battle/message.s"
+}
diff --git a/src/dakuten.s b/src/dakuten.s
index 756dafc..30c24cc 100755
--- a/src/dakuten.s
+++ b/src/dakuten.s
@@ -1,59 +1,64 @@
 """
+.include "bank20.i"
+
 Dakuten composite-glyph table + `lookup_dakuten` helper that maps a (prev_char, current_char) pair to the
 dakuten/handakuten composite tile pair.
 """
 
 
-.incbin "assets/dakuten.bin"
 
-lookup_dakuten:
-"""
-input: A 8bit: current char
-output: A 16bits: the resolved char
-"""
-{
-    php
-    sep #0x20
-    cmp.l assets_dakuten_bin
-    bmi _char_out_of_range
-    cmp.l assets_dakuten_bin + 2
-    bpl _char_out_of_range
-
-    phx
-    rep #0x20
-    and.w #0x00ff
-    sec
-    sbc.l assets_dakuten_bin
-    asl
-    tax
-
-    lda.l assets_dakuten_bin + 4, x
-
-    sep #0x20
-    plx
-    bra _exit
-_char_out_of_range:
-    xba
-    lda #0xff
-
-_exit:
-    plp
-    rtl
-}
+.alloc dakuten_block in bank20_reloc {
+    .incbin "assets/dakuten.bin"
+
+    lookup_dakuten:
+    """
+    input: A 8bit: current char
+    output: A 16bits: the resolved char
+    """
+    {
+        php
+        sep #0x20
+        cmp.l assets_dakuten_bin
+        bmi _char_out_of_range
+        cmp.l assets_dakuten_bin + 2
+        bpl _char_out_of_range
+
+        phx
+        rep #0x20
+        and.w #0x00ff
+        sec
+        sbc.l assets_dakuten_bin
+        asl
+        tax
+
+        lda.l assets_dakuten_bin + 4, x
+
+        sep #0x20
+        plx
+        bra _exit
+    _char_out_of_range:
+        xba
+        lda #0xff
+
+    _exit:
+        plp
+        rtl
+    }
+
+    _store_char_with_dakuten:
+    {
+        jsr.l lookup_dakuten
+        cmp #0xff
+        beq _skip
+        sta.l 0x7E0000, x
+    _skip:
+
+        sta.l 0x7E0040, x
 
-_store_char_with_dakuten:
-{
-    jsr.l lookup_dakuten
-    cmp #0xff
-    beq _skip
-    sta.l 0x7E0000, x
-_skip:
-
-    sta.l 0x7E0040, x
-
-;0187A9  9F 00 00 7E    STA $7E0000,X
-;0187AD  E8             INX
-;0187AE  E8             INX
-;0187AF  C8             INY
-    rtl
+    ;0187A9  9F 00 00 7E    STA $7E0000,X
+    ;0187AD  E8             INX
+    ;0187AE  E8             INX
+    ;0187AF  C8             INY
+        rtl
+    }
 }
diff --git a/src/dialog.s b/src/dialog.s
index 6c4544c..85dadc5 100644
--- a/src/dialog.s
+++ b/src/dialog.s
@@ -3,199 +3,206 @@ Dialog-text plumbing: 24-bit pointer resolver for translated bank 1-1/1-2/2 stri
 parser to load the next message.
 """
 .include "src/vwf.i"
+.include "bank20.i"
 
 .import "assets"
 .extern dialog_bank_ptr_base
 
-get_bank1_1_pointer:
-"""Get a 24-bit dialog pointer for bank 1-1."""
-    rep #0x20
-    lda.l assets_bank1_1_ptr, x
-    sta.b dialog_ptr
-    lda.w #0x0000
-    sep #0x20
-    lda.l assets_bank1_1_ptr + 2, x
-    sta.b dialog_ptr + 2
-    lda.b #0x01
-    rtl
-
-get_bank1_2_pointer:
-"""
-Get a 24-bit dialog pointer for bank 1-2.
-
-> Note: bank 1-1 is only 0x100 pointers long, not 0x200 as the text dump suggests.
-"""
-
-    rep #0x20
-    lda.l assets_bank1_1_ptr + 0x300, x
-    sta.b dialog_ptr
-    lda.w #0x0000
-    sep #0x20
-    lda.l assets_bank1_1_ptr + 0x300 + 2, x
-    sta.b dialog_ptr + 2
-    lda #0x01
-    rtl
-
-get_bank3_pointer:
-"""
-Compute pointer for NPC dialogs.
-
-Organized per room  ; a linear lookup inside the room finds the start of the string.
-"""
-
-    rep #0x20
-    lda.l dialog_bank_ptr_base + 0x600, x
-    sta.b dialog_ptr
-    lda.w #0x0000
-    sep #0x20
-    lda.l dialog_bank_ptr_base + 0x600 + 2, x
-    sta.b dialog_ptr + 2
-    lda #0x02
-    rtl
-
-compute_dialog_text_offset:
-"""Compute index into dialog pointer table from current text id ($B2)."""
-    lda.b 0xB2
-    sta.b dialog_ptr
-    stz.b dialog_ptr + 1
-    rep #0x20
-    lda.b dialog_ptr
-    clc
-    asl
-    adc.b dialog_ptr
-    tax
-    sep #0x20
-    rtl
-
-get_bank2_pointer:
-"""
-Get a 24-bit dialog pointer for bank 2.
-
-Walks the string character-by-character to handle variable-length encoding.
-"""
 
-{
-    rep #0x20
-    lda.b dialog_ptr
-    asl
-    clc
-    adc.b dialog_ptr
-    tax
-    lda.l assets_bank2_ptr, x
-    sta.b dialog_ptr
-    lda.w #0x0000
-    sep #0x20
-    lda.l assets_bank2_ptr + 2, x
-    sta.b dialog_ptr + 2
-    ldx.b dialog_ptr
-    lda.b 0xB2
-    beq _end
-    tay
-
-_loop:
-    jsr.w _load_letter_inc
-    bne _loop
-    jsr.w _load_letter_dec
-    pha
-    jsr.w _load_letter_inc
-    pla
-    cmp #0x03
-    beq _loop
-    pha
-    pla
-    cmp #0x04
-    beq _loop
-    cmp #0xfe
-    beq _loop
-    dey
-    bne _loop
-    inx
-
-_end:
-    stx.w 0x0772
-    stz.b 0xDD
-    rtl
-
-_load_letter_dec:
-    ldx.b dialog_ptr
-    dex
-    bmi _ok
-    dec.b dialog_ptr + 2
-    ldx.w #0xFFFF
-    bra _ok
-
-_load_letter_inc:
-    ldx.b dialog_ptr
-    inx
-    bmi _ok
-    inc.b dialog_ptr + 2
-    ldx.w #0x8000
-
-_ok:
-    stx.b dialog_ptr
-
-_load_letter:
-    ldx.b dialog_ptr
-    phb
-    lda.b dialog_ptr + 2
-    pha
-    plb
-    lda.w 0x0000, x
-    plb
-    pha
-    pla
-    rts
+.alloc dialog_block in bank20_reloc {
+    get_bank1_1_pointer:
+    """Get a 24-bit dialog pointer for bank 1-1."""
+        rep #0x20
+        lda.l assets_bank1_1_ptr, x
+        sta.b dialog_ptr
+        lda.w #0x0000
+        sep #0x20
+        lda.l assets_bank1_1_ptr + 2, x
+        sta.b dialog_ptr + 2
+        lda.b #0x01
+        rtl
+
+    get_bank1_2_pointer:
+    """
+    Get a 24-bit dialog pointer for bank 1-2.
+
+    > Note: bank 1-1 is only 0x100 pointers long, not 0x200 as the text dump suggests.
+    """
+
+
+        rep #0x20
+        lda.l assets_bank1_1_ptr + 0x300, x
+        sta.b dialog_ptr
+        lda.w #0x0000
+        sep #0x20
+        lda.l assets_bank1_1_ptr + 0x300 + 2, x
+        sta.b dialog_ptr + 2
+        lda #0x01
+        rtl
+
+    get_bank3_pointer:
+    """
+    Compute pointer for NPC dialogs.
+
+    Organized per room  ; a linear lookup inside the room finds the start of the string.
+    """
+
+
+        rep #0x20
+        lda.l dialog_bank_ptr_base + 0x600, x
+        sta.b dialog_ptr
+        lda.w #0x0000
+        sep #0x20
+        lda.l dialog_bank_ptr_base + 0x600 + 2, x
+        sta.b dialog_ptr + 2
+        lda #0x02
+        rtl
+
+    compute_dialog_text_offset:
+    """Compute index into dialog pointer table from current text id ($B2)."""
+        lda.b 0xB2
+        sta.b dialog_ptr
+        stz.b dialog_ptr + 1
+        rep #0x20
+        lda.b dialog_ptr
+        clc
+        asl
+        adc.b dialog_ptr
+        tax
+        sep #0x20
+        rtl
+
+    get_bank2_pointer:
+    """
+    Get a 24-bit dialog pointer for bank 2.
+
+    Walks the string character-by-character to handle variable-length encoding.
+    """
+
+
+    {
+        rep #0x20
+        lda.b dialog_ptr
+        asl
+        clc
+        adc.b dialog_ptr
+        tax
+        lda.l assets_bank2_ptr, x
+        sta.b dialog_ptr
+        lda.w #0x0000
+        sep #0x20
+        lda.l assets_bank2_ptr + 2, x
+        sta.b dialog_ptr + 2
+        ldx.b dialog_ptr
+        lda.b 0xB2
+        beq _end
+        tay
+
+    _loop:
+        jsr.w _load_letter_inc
+        bne _loop
+        jsr.w _load_letter_dec
+        pha
+        jsr.w _load_letter_inc
+        pla
+        cmp #0x03
+        beq _loop
+        pha
+        pla
+        cmp #0x04
+        beq _loop
+        cmp #0xfe
+        beq _loop
+        dey
+        bne _loop
+        inx
+
+    _end:
+        stx.w 0x0772
+        stz.b 0xDD
+        rtl
+
+    _load_letter_dec:
+        ldx.b dialog_ptr
+        dex
+        bmi _ok
+        dec.b dialog_ptr + 2
+        ldx.w #0xFFFF
+        bra _ok
+
+    _load_letter_inc:
+        ldx.b dialog_ptr
+        inx
+        bmi _ok
+        inc.b dialog_ptr + 2
+        ldx.w #0x8000
+
+    _ok:
+        stx.b dialog_ptr
+
+    _load_letter:
+        ldx.b dialog_ptr
+        phb
+        lda.b dialog_ptr + 2
+        pha
+        plb
+        lda.w 0x0000, x
+        plb
+        pha
+        pla
+        rts
+    }
+
+
+    _incpointer:
+    {
+        phx
+        ldx.w 0x0772
+        inx
+        bne _no_overflow
+        inc.b dialog_ptr + 2
+        ldx.w #0x8000
+
+    _no_overflow:
+        stx.w 0x0772
+        plx
+        rts
+    }
+
+
+    load_letter_inc:
+    """Advance dialog cursor by one character."""
+    {
+        ldx.w 0x0772
+        inx
+        cpx.w #0x0000
+        bne _no_overflow
+        inc.b dialog_ptr + 2
+        ldx.w #0x8000
+
+    _no_overflow:
+        stx.w 0x0772
+    }
+
+
+    load_letter:
+    """Peek the current character from the dialog stream into CURRENT_C."""
+        ldx.w 0x0772
+        phb
+        lda.b dialog_ptr + 2
+        pha
+        plb
+        lda.b #0x00
+        xba
+        lda.b #0x00
+        rep #0x20
+        lda.w 0x0000, x
+        sta.b CURRENT_C
+
+        sep #0x20
+        plb
+        pha
+        pla
+
+        rts
 }
-
-
-_incpointer:
-{
-    phx
-    ldx.w 0x0772
-    inx
-    bne _no_overflow
-    inc.b dialog_ptr + 2
-    ldx.w #0x8000
-
-_no_overflow:
-    stx.w 0x0772
-    plx
-    rts
-}
-
-
-load_letter_inc:
-"""Advance dialog cursor by one character."""
-{
-    ldx.w 0x0772
-    inx
-    cpx.w #0x0000
-    bne _no_overflow
-    inc.b dialog_ptr + 2
-    ldx.w #0x8000
-
-_no_overflow:
-    stx.w 0x0772
-}
-
-
-load_letter:
-"""Peek the current character from the dialog stream into CURRENT_C."""
-    ldx.w 0x0772
-    phb
-    lda.b dialog_ptr + 2
-    pha
-    plb
-    lda.b #0x00
-    xba
-    lda.b #0x00
-    rep #0x20
-    lda.w 0x0000, x
-    sta.b CURRENT_C
-
-    sep #0x20
-    plb
-    pha
-    pla
-
-    rts
diff --git a/src/ingame/drops_rolling.s b/src/ingame/drops_rolling.s
index 3e434c3..50769f1 100644
--- a/src/ingame/drops_rolling.s
+++ b/src/ingame/drops_rolling.s
@@ -47,7 +47,11 @@ DROPS_SCROLL_LIMIT := 3
 DROPS_SCROLL_PIXELS_PER_FRAME := 8
 DROPS_SCROLL_TOTAL_PIXELS := 16
 
-drops_rolling := 0x1BE0
+; Drops rolling state moved out of $1B00-$1BFF (vanilla menu / sprite
+; code writes to bytes past $1BEB) to clean $7E:9C30. Engine path needs
+; the full 35-byte struct ; the macro path only ever touched the first
+; 12 bytes so the original $1BE0 base worked there.
+drops_rolling := 0x9C30
 drops_rolling_top_row := drops_rolling + RollingBufferState.top_row
 drops_rolling_buffer_pos := drops_rolling + RollingBufferState.buffer_pos
 drops_rolling_edge_row := drops_rolling + RollingBufferState.edge_row
@@ -65,7 +69,7 @@ drops_hdma_copy_pending := drops_rolling + RollingBufferState.hdma_copy_pending
 ; doesn't collide with the engine's RollingBufferState fields. Other
 ; profiles read scroll_pos from a original menu byte ($1B1A field /
 ; $1BB7 treasure inventory); drops has no original equivalent.
-drops_scroll_pos := 0x1BEF
+drops_scroll_pos := 0x9C5F
 
 ; HDMA channel 4 (free in original treasure: enabled mask is $AD =
 ; ch7|ch5|ch3|ch2|ch0). Treasure inventory took ch6.
@@ -93,267 +97,478 @@ DROPS_SCROLL_STATE_SCROLLING := 1
 ; --- Profile hooks (stubs, real implementations land alongside the
 ;     drops geometry + tilemap layout work) -----------------------------------
 
-drops_ensure_hdma_initialized:
-"""Lazy init: pin BG4VOFS shadow to 0 + configure ch4 driving BG4VOFS on first scroll."""
-    rep #0x20
-    lda.w drops_rolling_base_scroll
-    cmp.w #0xFFFF
-    bne _drops_hdma_already_init
-
-; treasure_drops_window is anchored at BG (0, 0): top border at BG
-; row 0 = BG line 0. The HDMA header band offsets BG4VOFS so the
-; window appears on screen at y=24 (matching where the original
-; TreasureItemsWindow at $01:E275 lived). Items render at staging
-; rows 4,6,8,10 — engine body math uses base_scroll=0 so screen
-; y=32..95 reads exactly those rows.
-    lda.w #0xFFE8
-    ; -24: drops band starts at screen y=24. Items render at staging
-    ; rows 2,4,6,8 (DrawItemName lays glyphs into the row pointed at by
-    ; $1d = $29 + $40, so slot 0 with Y=$44 lands at row 2, not row 1).
-    ; Engine body math `scroll = base + slot*16 - row*16` collapses to
-    ; base for visible row 0..3 with buffer_pos=0; VOFS=-24 maps screen
-    ; y=32..95 to BG lines 8..71. Item 0 at BG row 2 = lines 16..23 lies
-    ; in the upper half of the body row 0 band (y=32..47).
-    sta.w drops_rolling_base_scroll
-
-; Configure HDMA channel 4: DIRECT mode 2 bytes, dest BG4VOFS ($2114),
-; src $7E:9880 (drops active table).
-    sep #0x20
-    lda #0x02
-    sta.l DROPS_HDMA4_CTRL
-    lda #0x14
-    sta.l DROPS_HDMA4_DEST
-    rep #0x20
-    lda.w #DROPS_HDMA_TABLE_ADDR
-    sta.l DROPS_HDMA4_SRC_LO
-    sep #0x20
-    lda #DROPS_HDMA_BANK
-    sta.l DROPS_HDMA4_SRC_BANK
-
-; Enable ch4 (BG4VOFS) only. TM HDMA mask via ch1 disabled — writes
-; to $212C per-scanline blank the screen for reasons not yet
-; understood (probably PPU/HDMA timing quirk in BGMODE 0). For now
-; rely on BG4 staging being zero-filled past the drops content (tile
-; 0 = transparent if CHR slot 0 is blank); residual bleed is the
-; trade-off.
-    lda.l 0x7E1BAE
-    ora #0x10
-    sta.l 0x7E1BAE
-    sta.w drops_hdma_enable
-    ; Build the initial HDMA shadow table + signal the NMI shadow→active
-    ; copy so ch4 has valid scroll values on the very first frame, not
-    ; just after the first scroll-trigger update.
-    jsr.w update_drops_scroll_hdma
-    rts
-
-_drops_hdma_already_init:
-    sep #0x20
-    rts
-
-drops_render_item_to_slot:
-"""Render one drops item from $7E:FF28 + edge_row*Item.__size into the BG3 buffer at $7E:D600 + slot_index*128 + 4."""
-    php
-    phb
-    lda #0x7E
-    pha
-    plb
-    rep #0x30
-    pha
-    phx
-    phy
-    lda.b 0x5a
-    pha
-    lda.b 0x29
-    pha
-    lda.b 0x45
-    pha
-    lda.b 0x33
-    pha
-    sep #0x20
-    lda.b 0x5d
-    pha
-    lda.b 0xDB
-    pha
-    rep #0x20
-    ; Pin $29 = $C600 (BG4 staging). TreasureItemsWindow is already
-    ; drawn on BG4 by original at $01:D817, so rendering drops items
-    ; INTO that same BG4 frame keeps the layout self-contained: one
-    ; window + items on one layer, one DMA path to VRAM.
-    lda.w #0xC600
-    sta.b 0x29
-    sep #0x20
-    lda.w drops_rolling_edge_row
-    asl
-    clc
-    adc #0x28
-    sta.b 0x5a
-    lda #0xFF
-    adc #0x00
-    sta.b 0x5b
-    rep #0x20
-    lda.b 0x5a
-    tax
-    sep #0x20
-    lda.l 0x7E0000 + Item.id, x
-    pha
-    lda.l 0x7E0000 + Item.qty, x
-    sta.b 0x5C
-    stz.b 0x34
-    pla
-    jsr.l check_can_use_item_trampoline
-    lda.w drops_rolling_slot_index
-    sta.b 0x5d
-    rep #0x20
-    lda.w drops_rolling_slot_index
-    and.w #0x00FF
-    xba
-    lsr
-    clc
-    adc.w #0x0044
-    tay
-    sep #0x20
-    jsr.l draw_item_slot_inner_trampoline
-    pla
-    sta.b 0xDB
-    pla
-    sta.b 0x5d
-    rep #0x20
-    pla
-    sta.b 0x33
-    pla
-    sta.b 0x45
-    pla
-    sta.b 0x29
-    pla
-    sta.b 0x5a
-    rep #0x10
-    ply
-    plx
-    pla
-    plb
-    plp
-    rts
-
-clear_drops_slot:
-"""Blank a single drops tilemap slot. STUB."""
-    rts
-
-update_drops_scroll_hdma:
-"""Build the drops HDMA shadow table via the shared engine."""
-    engine_update_scroll_hdma(drops_rolling, DROPS_HDMA_SHADOW, DROPS_BUFFER_SLOTS, DROPS_VISIBLE_ITEMS, _drops_hdma_header, _drops_hdma_footer, _drops_hdma_signal)  ; noqa: E501
-
-_drops_hdma_header:
-"""Header: 32 lines at base (-24) — off-screen 24 lines + 8 lines top border."""
-    sep #0x20
-    lda #32
-    sta.l DROPS_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w drops_rolling_base_scroll
-    sta.l DROPS_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_drops_hdma_footer:
-"""
-Footer: 25 lines bottom border + 87 lines trail keeping BG4 off content.
-
-Header (32) + body (5 visible × 16 = 80) + footer (25 + 87 = 112) = 224 visible scanlines.
-
-The border zone runs 25 lines (vs the natural 16) to step the trail's starting BG line past
-the bottom-border row. With the border zone @ -8, scanlines y=112..136 read BG rows 13..15
-(border + 2 blank padding rows). The trail @ -24 then starts at y=137, reading BG row 14
-— past the border row 13 — so the border tile doesn't get repeated for several scanlines as
-the trail kicks in.
-
-The +1 odd count (25 vs even) shifts the trail boundary to y=137 instead of y=136. ch6
-(treasure BG3VOFS) has a band boundary at y=136 (border-to-body)  ; the +1 keeps ch4 (drops)
-and ch6 from reloading on the same scanline.
-"""
-
-
-    sep #0x20
-    lda #25
-    sta.l DROPS_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w #0xFFF8
-    ; -8: bottom border (BG row 13) at screen y=112..137. 25 lines covers BG rows 13..15
-    ; (border + 2 blank padding) so the trail @ -24 starts at y=137 reading BG row 14
-    ; instead of looping back to row 13 — prevents the border tile repeating visually.
-    sta.l DROPS_HDMA_SHADOW, x
-    inx
-    inx
-    sep #0x20
-    lda #87
-    sta.l DROPS_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w #0xFFE8
-    ; -24 again: y=137..223 reads BG rows 14..24 (empty padding) so BG4 stays off content.
-    sta.l DROPS_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_drops_hdma_signal:
-"""NMI shadow-copy signal for the drops ch4 table."""
-    sep #0x20
-    lda #0x01
-    sta.w drops_hdma_copy_pending
-    rts
-
-
-; --- Engine instantiations -------------------------------------------------
-
-drops_init_impl:
-"""
-Init drops rolling buffer. Drops render onto BG4 staging ($7E:C600) alongside TreasureItemsWindow
-($01:E275 def_window 1, 3, 27, 10) which the draw_window hook redraws first so the engine can write items into
-the just-drawn frame without the original $01:D817 DrawWindow call clobbering them.
-"""
-    engine_init_rolling_buffer(drops_rolling, DROPS_VISIBLE_ITEMS, _drops_draw_window, drops_ensure_hdma_initialized, drops_render_item_to_slot)  ; noqa: E501
-
-
-_drops_draw_window:
-"""Pre-render hook: SelectBG4 + DrawWindow(treasure_drops_window) so $29=$C600 is live before items render."""
-    sep #0x20
-    jsr.l _drops_select_bg4_trampoline
-    rep #0x10
-    ldy.w #treasure_drops_window
-    jsr.l draw_window_trampoline
-    sep #0x10
-    rts
-
-drops_scroll_down_prepare:
-"""Drops profile scroll-down pre-render."""
-    engine_scroll_down_prepare(drops_rolling, drops_scroll_pos, DROPS_SCROLL_LIMIT, DROPS_VISIBLE_ITEMS, DROPS_BUFFER_SLOTS, drops_ensure_hdma_initialized, drops_render_item_to_slot, update_drops_scroll_hdma)  ; noqa: E501
-
-drops_scroll_up_prepare:
-"""Drops profile scroll-up pre-render."""
-    engine_scroll_up_prepare(drops_rolling, drops_scroll_pos, DROPS_BUFFER_SLOTS, drops_ensure_hdma_initialized, drops_render_item_to_slot, update_drops_scroll_hdma)  ; noqa: E501
-
-drops_start_scroll_down_impl:
-"""Drops profile: kick scroll-down state machine."""
-    engine_start_scroll_down(drops_rolling, drops_scroll_pos, DROPS_VISIBLE_ITEMS, DROPS_BUFFER_SLOTS, DROPS_SCROLL_TOTAL_PIXELS, DROPS_SCROLL_PIXELS_PER_FRAME, drops_ensure_hdma_initialized, drops_render_item_to_slot, update_drops_scroll_hdma)  ; noqa: E501
-
-drops_start_scroll_up_impl:
-"""Drops profile: kick scroll-up state machine."""
-    engine_start_scroll_up(drops_rolling, drops_scroll_pos, DROPS_BUFFER_SLOTS, DROPS_SCROLL_TOTAL_PIXELS, drops_ensure_hdma_initialized, drops_render_item_to_slot, update_drops_scroll_hdma)  ; noqa: E501
-
-drops_update_scroll_frame_impl:
-"""Drops profile: per-frame scroll animation tick."""
-    engine_update_scroll_frame(drops_rolling, DROPS_SCROLL_PIXELS_PER_FRAME, update_drops_scroll_hdma)
-
-drops_finish_scroll_impl:
-"""Drops profile: end-of-animation pre-render + cleanup."""
-    engine_finish_scroll(drops_rolling, drops_scroll_pos, DROPS_VISIBLE_ITEMS, DROPS_BUFFER_SLOTS, DROPS_TOTAL_ITEMS, drops_render_item_to_slot, update_drops_scroll_hdma)  ; noqa: E501
-
-drops_refresh_slots_impl:
-"""Drops profile: re-render all 5 slots (original redraw helper path)."""
-    engine_refresh_slots(drops_rolling, drops_scroll_pos, DROPS_BUFFER_SLOTS, drops_render_item_to_slot)
-
-drops_swap_redraw_impl:
-"""Drops profile: post-swap re-render."""
-    engine_swap_redraw(drops_rolling, drops_scroll_pos, DROPS_BUFFER_SLOTS, DROPS_TOTAL_ITEMS, drops_ensure_hdma_initialized, drops_render_item_to_slot, clear_drops_slot, update_drops_scroll_hdma)  ; noqa: E501
+.include "../bank20.i"
+
+.alloc drops_rolling_block in bank20_reloc {
+    drops_ensure_hdma_initialized:
+    """Lazy init: pin BG4VOFS shadow to 0 + configure ch4 driving BG4VOFS on first scroll."""
+        rep #0x20
+        lda.w drops_rolling_base_scroll
+        cmp.w #0xFFFF
+        bne _drops_hdma_already_init
+
+    ; treasure_drops_window is anchored at BG (0, 0): top border at BG
+    ; row 0 = BG line 0. The HDMA header band offsets BG4VOFS so the
+    ; window appears on screen at y=24 (matching where the original
+    ; TreasureItemsWindow at $01:E275 lived). Items render at staging
+    ; rows 4,6,8,10 — engine body math uses base_scroll=0 so screen
+    ; y=32..95 reads exactly those rows.
+        lda.w #0xFFE8
+        ; -24: drops band starts at screen y=24. Items render at staging
+        ; rows 2,4,6,8 (DrawItemName lays glyphs into the row pointed at by
+        ; $1d = $29 + $40, so slot 0 with Y=$44 lands at row 2, not row 1).
+        ; Engine body math `scroll = base + slot*16 - row*16` collapses to
+        ; base for visible row 0..3 with buffer_pos=0; VOFS=-24 maps screen
+        ; y=32..95 to BG lines 8..71. Item 0 at BG row 2 = lines 16..23 lies
+        ; in the upper half of the body row 0 band (y=32..47).
+        sta.w drops_rolling_base_scroll
+
+    ; Configure HDMA channel 4: DIRECT mode 2 bytes, dest BG4VOFS ($2114),
+    ; src $7E:9880 (drops active table).
+        sep #0x20
+        lda #0x02
+        sta.l DROPS_HDMA4_CTRL
+        lda #0x14
+        sta.l DROPS_HDMA4_DEST
+        rep #0x20
+        lda.w #DROPS_HDMA_TABLE_ADDR
+        sta.l DROPS_HDMA4_SRC_LO
+        sep #0x20
+        lda #DROPS_HDMA_BANK
+        sta.l DROPS_HDMA4_SRC_BANK
+
+    ; Enable ch4 (BG4VOFS) only. TM HDMA mask via ch1 disabled — writes
+    ; to $212C per-scanline blank the screen for reasons not yet
+    ; understood (probably PPU/HDMA timing quirk in BGMODE 0). For now
+    ; rely on BG4 staging being zero-filled past the drops content (tile
+    ; 0 = transparent if CHR slot 0 is blank); residual bleed is the
+    ; trade-off.
+        lda.l 0x7E1BAE
+        ora #0x10
+        sta.l 0x7E1BAE
+        sta.w drops_hdma_enable
+        ; Build the initial HDMA shadow table + signal the NMI shadow→active
+        ; copy so ch4 has valid scroll values on the very first frame, not
+        ; just after the first scroll-trigger update.
+        jsr.w update_drops_scroll_hdma
+        rts
+
+    _drops_hdma_already_init:
+        sep #0x20
+        rts
+
+    drops_render_item_to_slot:
+    """Render one drops item from $7E:FF28 + edge_row*Item.__size into the BG3 buffer at $7E:D600 + slot_index*128 + 4."""
+        php
+        phb
+        lda #0x7E
+        pha
+        plb
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x5a
+        pha
+        lda.b 0x29
+        pha
+        lda.b 0x45
+        pha
+        lda.b 0x33
+        pha
+        sep #0x20
+        lda.b 0x5d
+        pha
+        lda.b 0xDB
+        pha
+        rep #0x20
+        ; Pin $29 = $C600 (BG4 staging). TreasureItemsWindow is already
+        ; drawn on BG4 by original at $01:D817, so rendering drops items
+        ; INTO that same BG4 frame keeps the layout self-contained: one
+        ; window + items on one layer, one DMA path to VRAM.
+        lda.w #0xC600
+        sta.b 0x29
+        sep #0x20
+        lda.w drops_rolling_edge_row
+        asl
+        clc
+        adc #0x28
+        sta.b 0x5a
+        lda #0xFF
+        adc #0x00
+        sta.b 0x5b
+        rep #0x20
+        lda.b 0x5a
+        tax
+        sep #0x20
+        lda.l 0x7E0000 + Item.id, x
+        pha
+        lda.l 0x7E0000 + Item.qty, x
+        sta.b 0x5C
+        stz.b 0x34
+        pla
+        jsr.l check_can_use_item_trampoline
+        ; --- VWF tile-id offset for drops (disjoint from treasure inventory) ---
+        ; `items_menu_vwf.draw_field_item_name` computes
+        ; `tile_id_base = FIELD_ITEM_VWF_TILE_BASE + $5D * K`. Treasure inventory
+        ; calls the same helper with $5D = treasure_slot_index ; if drops were
+        ; to pass its raw slot_index too, both panels would land on the same
+        ; tile_ids in BG3 CHR ($5000..$56E0) and the second renderer of each
+        ; frame would overwrite the first one's glyphs (the garbage-tiles bug
+        ; visible in the chest-with-monster-drop screen). Offset by 11 so
+        ; drops slots 0..5 land at tile_ids $16E..$1A9 (CHR $56E0..$5A90),
+        ; past treasure's 6-buffer-slot region at $100..$13B. drops's tilemap
+        ; target stays at $7E:C600 (its own BG3 staging surface) so the offset
+        ; only affects CHR allocation, not where the glyphs land on screen.
+        ;
+        ; VWF_CALLER_CTX=1 also tells items_menu_vwf to write the SECONDARY
+        ; flush descriptor (DROPS_VWF_VRAM_DEST_WORD + DROPS_VWF_BYTE_COUNT +
+        ; DROPS_VWF_CHR_SRC_OFFSET) and redirect VWF_CHR_DIRTY -> DIRTY_B so
+        ; the NMI flush hits drops's VRAM range without trampling treasure's
+        ; primary descriptor. Cleared after the render so subsequent
+        ; treasure-side calls fall back to primary.
+        lda.w drops_rolling_slot_index
+        clc
+        adc #DROPS_VWF_TILE_SLOT_OFFSET
+        sta.b 0x5d
+        lda #0x01
+        sta.l VWF_CALLER_CTX
+        rep #0x20
+        lda.w drops_rolling_slot_index
+        and.w #0x00FF
+        xba
+        lsr
+        clc
+        adc.w #0x0044
+        tay
+        sep #0x20
+        jsr.l draw_item_slot_inner_trampoline
+        lda #0x00
+        sta.l VWF_CALLER_CTX
+        pla
+        sta.b 0xDB
+        pla
+        sta.b 0x5d
+        rep #0x20
+        pla
+        sta.b 0x33
+        pla
+        sta.b 0x45
+        pla
+        sta.b 0x29
+        pla
+        sta.b 0x5a
+        rep #0x10
+        ply
+        plx
+        pla
+        plb
+        plp
+        rts
+
+    clear_drops_slot:
+    """Blank a single drops tilemap slot. STUB."""
+        rts
+
+    update_drops_scroll_hdma:
+    """Build the drops HDMA shadow table via the shared engine."""
+        ; Build drops HDMA scroll table. Inlined from the former
+        ; engine_update_scroll_hdma macro for the same reason as treasure.
+    {
+        php
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x40
+        pha
+        lda.b 0x42
+        pha
+        ldx.w #0x0000
+        jsr.w _drops_hdma_header
+        stz.b 0x42
+
+    _row_loop:
+        lda.w drops_rolling + RollingBufferState.buffer_pos
+        and.w #0x00FF
+        clc
+        adc.b 0x42
+
+    _mod_loop:
+        cmp.w #DROPS_BUFFER_SLOTS
+        bcc _mod_done
+        sec
+        sbc.w #DROPS_BUFFER_SLOTS
+        bra _mod_loop
+
+    _mod_done:
+        asl
+        asl
+        asl
+        asl
+        sta.b 0x40
+        lda.b 0x42
+        and.w #0x00FF
+        asl
+        asl
+        asl
+        asl
+        eor.w #0xFFFF
+        inc
+        clc
+        adc.b 0x40
+        clc
+        adc.w drops_rolling + RollingBufferState.base_scroll
+        sta.b 0x40
+        sep #0x20
+        lda #16
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.b 0x40
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        inx
+        rep #0x20
+        inc.b 0x42
+        lda.b 0x42
+        cmp.w #DROPS_VISIBLE_ITEMS
+        bcs _row_loop_done
+        jmp.w _row_loop
+
+    _row_loop_done:
+        jsr.w _drops_hdma_footer
+        sep #0x20
+        lda #0x00
+        sta.l DROPS_HDMA_SHADOW, x
+        jsr.w _drops_hdma_signal
+        rep #0x20
+        pla
+        sta.b 0x42
+        pla
+        sta.b 0x40
+        ply
+        plx
+        pla
+        plp
+        rts
+    }
+
+    _drops_hdma_header:
+    """Header: 32 lines at base (-24) — off-screen 24 lines + 8 lines top border."""
+        sep #0x20
+        lda #32
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w drops_rolling_base_scroll
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _drops_hdma_footer:
+    """
+    Footer: 25 lines bottom border + 87 lines trail keeping BG4 off content.
+
+    Header (32) + body (5 visible × 16 = 80) + footer (25 + 87 = 112) = 224 visible scanlines.
+
+    The border zone runs 25 lines (vs the natural 16) to step the trail's starting BG line past
+    the bottom-border row. With the border zone @ -8, scanlines y=112..136 read BG rows 13..15
+    (border + 2 blank padding rows). The trail @ -24 then starts at y=137, reading BG row 14
+    — past the border row 13 — so the border tile doesn't get repeated for several scanlines as
+    the trail kicks in.
+
+    The +1 odd count (25 vs even) shifts the trail boundary to y=137 instead of y=136. ch6
+    (treasure BG3VOFS) has a band boundary at y=136 (border-to-body)  ; the +1 keeps ch4 (drops)
+    and ch6 from reloading on the same scanline.
+    """
+
+
+        sep #0x20
+        lda #25
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w #0xFFF8
+        ; -8: bottom border (BG row 13) at screen y=112..137. 25 lines covers BG rows 13..15
+        ; (border + 2 blank padding) so the trail @ -24 starts at y=137 reading BG row 14
+        ; instead of looping back to row 13 — prevents the border tile repeating visually.
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        inx
+        sep #0x20
+        lda #87
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w #0xFFE8
+        ; -24 again: y=137..223 reads BG rows 14..24 (empty padding) so BG4 stays off content.
+        sta.l DROPS_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _drops_hdma_signal:
+    """NMI shadow-copy signal for the drops ch4 table."""
+        sep #0x20
+        lda #0x01
+        sta.w drops_hdma_copy_pending
+        rts
+
+
+    ; --- Engine instantiations -------------------------------------------------
+
+    drops_init_impl:
+    """
+    Init drops rolling buffer via the bank-20 engine. State + hook
+    far-ptrs live at $7E:9C30 (relocated out of $1B00-$1BFF). Drops render
+    onto BG4 staging ($7E:C600) alongside TreasureItemsWindow ($01:E275
+    def_window 1, 3, 27, 10) which the draw_window hook redraws first so
+    the engine can write items into the just-drawn frame without the
+    original $01:D817 DrawWindow call clobbering them.
+    """
+        php
+        rep #0x30
+        sep #0x20
+        lda.b #DROPS_VISIBLE_ITEMS
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.visible_rows
+        lda.b #0x02
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.slot_height_tiles
+        ; item_list_ptr = $7E:FF28 (treasure drops table)
+        lda.b #0x28
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.item_list_ptr
+        lda.b #0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.item_list_ptr + 1
+        lda.b #0x7E
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.item_list_ptr + 2
+        lda.b #DROPS_TOTAL_ITEMS
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.item_count
+        lda.b #0x04
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.hdma_channel
+        lda.b #0x80
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.vwf_cfg_ptr
+        lda.b #0x70
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.vwf_cfg_ptr + 1
+        lda.b #0x70
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.vwf_cfg_ptr + 2
+        lda.b #drops_fn_render_slot_trampoline & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_render_slot
+        lda.b #( drops_fn_render_slot_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_render_slot + 1
+        lda.b #( drops_fn_render_slot_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_render_slot + 2
+        lda.b #drops_fn_update_hdma_trampoline & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_update_hdma
+        lda.b #( drops_fn_update_hdma_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_update_hdma + 1
+        lda.b #( drops_fn_update_hdma_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_update_hdma + 2
+        lda.b #drops_fn_draw_window_trampoline & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_draw_window
+        lda.b #( drops_fn_draw_window_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_draw_window + 1
+        lda.b #( drops_fn_draw_window_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.fn_draw_window + 2
+        lda.b #ROLLING_MENU_ID_DROPS
+        sta.l 0x7E0000 + drops_rolling + RollingBufferState.menu_id
+        plp
+        php
+        rep #0x10
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_init
+        plp
+        rtl
+
+    drops_fn_render_slot_trampoline:
+    """Bank-20 RTL wrapper around `drops_render_item_to_slot`."""
+        php
+        jsr.w drops_render_item_to_slot
+        plp
+        rtl
+
+    drops_fn_update_hdma_trampoline:
+    """Bank-20 RTL wrapper around `drops_ensure_hdma_initialized`."""
+        php
+        jsr.w drops_ensure_hdma_initialized
+        plp
+        rtl
+
+    drops_fn_draw_window_trampoline:
+    """Bank-20 RTL wrapper around `_drops_draw_window`."""
+        php
+        jsr.w _drops_draw_window
+        plp
+        rtl
+
+
+    _drops_draw_window:
+    """Pre-render hook: SelectBG4 + DrawWindow(treasure_drops_window) so $29=$C600 is live before items render."""
+        sep #0x20
+        jsr.l _drops_select_bg4_trampoline
+        rep #0x10
+        ldy.w #treasure_drops_window
+        jsr.l draw_window_trampoline
+        sep #0x10
+        rts
+
+    drops_start_scroll_down_impl:
+    """Drops profile: kick scroll-down state machine via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + drops_scroll_pos
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_down
+        plp
+        rtl
+
+    drops_start_scroll_up_impl:
+    """Drops profile: kick scroll-up state machine via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + drops_scroll_pos
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_up
+        plp
+        rtl
+
+    drops_update_scroll_frame_impl:
+    """Drops profile: per-frame scroll animation tick."""
+        php
+        rep #0x10
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_update_scroll_frame
+        plp
+        rtl
+
+    drops_finish_scroll_impl:
+    """Drops profile: end-of-animation cleanup via the bank-20 engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + drops_scroll_pos
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_finish_scroll
+        plp
+        rtl
+
+    drops_refresh_slots_impl:
+    """Drops profile: re-render all 6 slots via the bank-20 engine."""
+        php
+        sep #0x20
+        rep #0x10
+        lda.l 0x7E0000 + drops_scroll_pos
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_refresh_slots
+        plp
+        rtl
+
+    drops_swap_redraw_impl:
+    """Drops profile: post-swap re-render via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + drops_scroll_pos
+        ldx.w #drops_rolling
+        jsr.l rolling_engine.rolling_engine_swap_redraw
+        plp
+        rtl
+}
diff --git a/src/ingame/free_space.s b/src/ingame/free_space.s
index 3b8d511..e5fd5f6 100644
--- a/src/ingame/free_space.s
+++ b/src/ingame/free_space.s
@@ -14,12 +14,13 @@ relocating other code.
 draw_window = 0x0180d9
 
 .alloc bank01_slack_pre_inventory in bank01_slack {
-
 check_if_description_was_rendered:
 """
 Skip redrawing an item description if its text pointer matches `render.last_drawn_text_ptr` and the
 auto-counter ($4218/$4219) is non-zero.
 """
+
+
     pha
     lda.l 0x004218
     ora.l 0x004219
@@ -54,8 +55,9 @@ Open a menu window at the cursor and render its VWF message  ; advances Y past t
 delegating to `_draw_vwf_message_pos`.
 """
 
+
     jsr.w draw_window
-    ; NOTE: quirks from the hardcore bank switching can be solved by loading the bank in A before the call.
+; NOTE: quirks from the hardcore bank switching can be solved by loading the bank in A before the call.
     pha
     rep #0x20
     tya
@@ -74,24 +76,26 @@ draw_vwf_message_pos_with_bank:
 Like `_draw_vwf_message_pos` but pre-loads the menu-strings bank into A so the trampoline can pick the right
 asset bank.
 """
+
+
     lda.b #messages.use_on_whom >> 16
 
 _draw_vwf_message_pos:
     jsr.l items_description.draw_trampoline_pos
     rts
 
-.if 0 {
+    .if 0 {
 transform_window_trampoline:
 """JML trampoline into `transform_window_far`."""
     jmp.l transform_window_far
-}
+    }
 
 copy_text_with_dakuten:
 """Near-call wrapper around `copy_text_with_dakuten_far` for callers in the same bank."""
     jsr.l copy_text_with_dakuten_far
     rts
 
-.if DEBUG {
+    .if DEBUG {
 display_build_number:
 """Render the `BUILD_DATE + version` string at column 1, row 27 of the title screen (DEBUG builds only)."""
     {
@@ -103,15 +107,16 @@ display_build_number:
     jsr.w 0x8798  ; copy text at position.
     rts
     }
+    }
 }
 
-}  ; end .alloc bank01_slack_pre_inventory
+; end .alloc bank01_slack_pre_inventory
 
 ; ============================================================================
 ; Inventory Rolling Buffer Trampolines and Handlers
 ; ============================================================================
 .alloc bank01_slack_inventory in bank01_slack {
-.if INVENTORY_ROLLING_BUFFER {
+    .if INVENTORY_ROLLING_BUFFER {
 swap_redraw_trampoline:
 """JML trampoline into `swap_redraw_hook_impl` for the inventory swap redraw path."""
     jsr.w swap_redraw_hook_impl
@@ -217,7 +222,8 @@ Re-renders all visible slots to show updated quantity or empty slot
 
     jsr.w swap_redraw_hook_impl
     rts
+    }
 }
 
 
-}  ; end .alloc bank01_slack_inventory
+; end .alloc bank01_slack_inventory
diff --git a/src/ingame/init_bg_scroll_hdma.s b/src/ingame/init_bg_scroll_hdma.s
index 3e1668f..d3f554e 100644
--- a/src/ingame/init_bg_scroll_hdma.s
+++ b/src/ingame/init_bg_scroll_hdma.s
@@ -1,244 +1,249 @@
 """Relocated `InitBGScrollHDMA` (from $01:EBD2): builds the BG2/BG3 V-scroll HDMA tables in WRAM."""
 
-init_bg_scroll_hdma:
-"""Init BG Scroll HDMA — relocated from $01:EBD2."""
-    tdc
-    tax
-
-_at_ebd4:
-    lda 0x0DFF29, x  ; load bg scroll hdma tables
-    sta 0x75FD, x
-    inx
-    cpx.w #0x0015
-    bne _at_ebd4
-    tdc
-    tax
-
-_at_ebe3:
-    sta 0x7612, x  ; clear hdma data
-    inx
-    cpx #0x1620
-    bne _at_ebe3
-    tdc
-    tax
-    lda #0xFE
-
-_at_ebf0:
-    sta 0x7D14, x  ; bg3 v-scroll
-    dec
-    inx
-    inx
-    inx
-    inx
-    cpx #0x0380
-    bne _at_ebf0
-    ldx #0x0230
-
-_at_ec00:
-    lda 0x7D14, x
-    dec
-    sta 0x7994, x  ; bg2 v-scroll
-    inx
-    inx
-    inx
-    inx
-    cpx #0x0280
-    bne _at_ec00
-    rep #0x20
-    tdc
-    tax
-    lda #0x0173
-    ldy.w #0x0008
-
-_at_ec1a:
-    sta 0x8094, x
-    pha
-    clc
-    adc.w #0x0068
-    sta 0x8314, x
-    clc
-    adc.w #0x00F0
-    sta 0x8AF4, x
-    pla
-    dey
-    bne _at_ec37
-    clc
-    adc.w #0x0004
-    ldy.w #0x000C
-
-_at_ec37:
-    cpx #0x0110
-    bne _at_ec40
-    clc
-    adc.w #0x0004
-
-_at_ec40:
-    inx
-    inx
-    inx
-    inx
-    cpx #0x0130
-    bne _at_ec1a
-    tdc
-    tax
-    lda #0x016F
-    ldy.w #0x0008
-
-_at_ec51:
-    inc 0x81D3, x
-    sta 0x81D4, x
-    dey
-    bne _at_ec61
-    clc
-    adc.w #0x0004
-    ldy.w #0x000C
-
-_at_ec61:
-    cpx #0x0110
-    bne _at_ec6a
-    clc
-    adc #0x0134
-
-_at_ec6a:
-    inx
-    inx
-    inx
-    inx
-    cpx #0x0130
-    bne _at_ec51
-    tdc
-    tax
-    lda.w #0x006B
-    ldy.w #0x0008
-
-_at_ec7b:
-    sta 0x8454, x
-    dey
-    bne _at_ec88
-    clc
-    adc.w #0x0004
-    ldy.w #0x000C
-
-_at_ec88:
-    cpx.w #0x0088
-    bne _at_ec91
-    clc
-    adc.w #0x0004
-
-_at_ec91:
-    inx
-    inx
-    inx
-    inx
-    cpx.w #0x00A0
-    bne _at_ec7b
-    tdc
-    tax
-
-_at_ec9c:
-    lda 0x8072, x
-    sta 0x81C2, x
-    sta 0x8442, x
-    inx
-    inx
-    cpx.w #0x0010
-    bne _at_ec9c
-    tdc
-    tax
-
-_at_ecae:
-    lda #0x0101
-    sta 0x84F2, x
-    inx
-    inx
-    inx
-    inx
-    cpx #0x0100
-    bne _at_ecae
-    tdc
-    tax
-    lda.w #0x0053
-    ldy.w #0x0008
-
-_at_ecc5:
-    sta 0x85F4, x
-    sta 0x8874, x
-    pha
-    sec
-    sbc.w #0x000C
-    sta 0x8674, x
-    sta 0x88F4, x
-    sec
-    sbc.w #0x000C
-    sta 0x86F4, x
-    sta 0x8974, x
-    sec
-    sbc.w #0x000C
-    sta 0x8774, x
-    sta 0x89F4, x
-    sec
-    sbc.w #0x000C
-    sta 0x87F4, x
-    sta 0x8A74, x
-    pla
-    dey
-    bne _at_ecfc
-    clc
-    adc.w #0x0004
-
-_at_ecfc:
-    pha
-    lda.w #0x00AC
-    sta 0x8872, x
-    sta 0x88F2, x
-    sta 0x8972, x
-    sta 0x89F2, x
-    sta 0x8A72, x
-    lda #0x01BC
-    sta 0x85F2, x
-    sta 0x8672, x
-    sta 0x86F2, x
-    sta 0x8772, x
-    sta 0x87F2, x
-    pla
-    inx
-    inx
-    inx
-    inx
-    cpx.w #0x0070
-    bne _at_ecc5
-    ldx.w #0x001C
-    ldy.w #0x0004
-    lda #0x0134
-
-_at_ed34:
-    sta 0x7D14, x
-    dey
-    bne _at_ed3e
-    clc
-    adc.w #0x0004
-
-_at_ed3e:
-    inx
-    inx
-    inx
-    inx
-    cpx.w #0x0080
-    bne _at_ed34
-    tdc
-    tax
-
-_at_ed49:
-    lda #0x0100
-    sta 0x8C32, x
-    lda #0x0160
-    sta 0x8C34, x
-    inx
-    inx
-    inx
-    inx
-    cpx.w #0x0080
-    bne _at_ed49
-    tdc
-    sep #0x20
-    rtl
+
+.include "../bank20.i"
+
+.alloc init_bg_scroll_hdma_block in bank20_reloc {
+    init_bg_scroll_hdma:
+    """Init BG Scroll HDMA — relocated from $01:EBD2."""
+        tdc
+        tax
+
+    _at_ebd4:
+        lda 0x0DFF29, x  ; load bg scroll hdma tables
+        sta 0x75FD, x
+        inx
+        cpx.w #0x0015
+        bne _at_ebd4
+        tdc
+        tax
+
+    _at_ebe3:
+        sta 0x7612, x  ; clear hdma data
+        inx
+        cpx #0x1620
+        bne _at_ebe3
+        tdc
+        tax
+        lda #0xFE
+
+    _at_ebf0:
+        sta 0x7D14, x  ; bg3 v-scroll
+        dec
+        inx
+        inx
+        inx
+        inx
+        cpx #0x0380
+        bne _at_ebf0
+        ldx #0x0230
+
+    _at_ec00:
+        lda 0x7D14, x
+        dec
+        sta 0x7994, x  ; bg2 v-scroll
+        inx
+        inx
+        inx
+        inx
+        cpx #0x0280
+        bne _at_ec00
+        rep #0x20
+        tdc
+        tax
+        lda #0x0173
+        ldy.w #0x0008
+
+    _at_ec1a:
+        sta 0x8094, x
+        pha
+        clc
+        adc.w #0x0068
+        sta 0x8314, x
+        clc
+        adc.w #0x00F0
+        sta 0x8AF4, x
+        pla
+        dey
+        bne _at_ec37
+        clc
+        adc.w #0x0004
+        ldy.w #0x000C
+
+    _at_ec37:
+        cpx #0x0110
+        bne _at_ec40
+        clc
+        adc.w #0x0004
+
+    _at_ec40:
+        inx
+        inx
+        inx
+        inx
+        cpx #0x0130
+        bne _at_ec1a
+        tdc
+        tax
+        lda #0x016F
+        ldy.w #0x0008
+
+    _at_ec51:
+        inc 0x81D3, x
+        sta 0x81D4, x
+        dey
+        bne _at_ec61
+        clc
+        adc.w #0x0004
+        ldy.w #0x000C
+
+    _at_ec61:
+        cpx #0x0110
+        bne _at_ec6a
+        clc
+        adc #0x0134
+
+    _at_ec6a:
+        inx
+        inx
+        inx
+        inx
+        cpx #0x0130
+        bne _at_ec51
+        tdc
+        tax
+        lda.w #0x006B
+        ldy.w #0x0008
+
+    _at_ec7b:
+        sta 0x8454, x
+        dey
+        bne _at_ec88
+        clc
+        adc.w #0x0004
+        ldy.w #0x000C
+
+    _at_ec88:
+        cpx.w #0x0088
+        bne _at_ec91
+        clc
+        adc.w #0x0004
+
+    _at_ec91:
+        inx
+        inx
+        inx
+        inx
+        cpx.w #0x00A0
+        bne _at_ec7b
+        tdc
+        tax
+
+    _at_ec9c:
+        lda 0x8072, x
+        sta 0x81C2, x
+        sta 0x8442, x
+        inx
+        inx
+        cpx.w #0x0010
+        bne _at_ec9c
+        tdc
+        tax
+
+    _at_ecae:
+        lda #0x0101
+        sta 0x84F2, x
+        inx
+        inx
+        inx
+        inx
+        cpx #0x0100
+        bne _at_ecae
+        tdc
+        tax
+        lda.w #0x0053
+        ldy.w #0x0008
+
+    _at_ecc5:
+        sta 0x85F4, x
+        sta 0x8874, x
+        pha
+        sec
+        sbc.w #0x000C
+        sta 0x8674, x
+        sta 0x88F4, x
+        sec
+        sbc.w #0x000C
+        sta 0x86F4, x
+        sta 0x8974, x
+        sec
+        sbc.w #0x000C
+        sta 0x8774, x
+        sta 0x89F4, x
+        sec
+        sbc.w #0x000C
+        sta 0x87F4, x
+        sta 0x8A74, x
+        pla
+        dey
+        bne _at_ecfc
+        clc
+        adc.w #0x0004
+
+    _at_ecfc:
+        pha
+        lda.w #0x00AC
+        sta 0x8872, x
+        sta 0x88F2, x
+        sta 0x8972, x
+        sta 0x89F2, x
+        sta 0x8A72, x
+        lda #0x01BC
+        sta 0x85F2, x
+        sta 0x8672, x
+        sta 0x86F2, x
+        sta 0x8772, x
+        sta 0x87F2, x
+        pla
+        inx
+        inx
+        inx
+        inx
+        cpx.w #0x0070
+        bne _at_ecc5
+        ldx.w #0x001C
+        ldy.w #0x0004
+        lda #0x0134
+
+    _at_ed34:
+        sta 0x7D14, x
+        dey
+        bne _at_ed3e
+        clc
+        adc.w #0x0004
+
+    _at_ed3e:
+        inx
+        inx
+        inx
+        inx
+        cpx.w #0x0080
+        bne _at_ed34
+        tdc
+        tax
+
+    _at_ed49:
+        lda #0x0100
+        sta 0x8C32, x
+        lda #0x0160
+        sta 0x8C34, x
+        inx
+        inx
+        inx
+        inx
+        cpx.w #0x0080
+        bne _at_ed49
+        tdc
+        sep #0x20
+        rtl
+}
diff --git a/src/ingame/inventory_rolling.s b/src/ingame/inventory_rolling.s
index 31546ad..6542d98 100644
--- a/src/ingame/inventory_rolling.s
+++ b/src/ingame/inventory_rolling.s
@@ -90,895 +90,1120 @@ HDMA5_SRC_BANK := 0x4354  ; Source bank
 HDMA5_IND_BANK := 0x4357  ; Indirect bank
 HDMAEN := 0x420C  ; HDMA enable register
 
-init_menu_inventory_hdma:
-"""
-Sets up HDMA channel 5 for per-scanline BG1 vertical scroll control
-Called when entering the item menu
-
-HDMA mode: $02 = write 2 bytes to same register, DIRECT mode (like FF6)
-Register: $210E (BG1VOFS)
-Table format: count_byte, lo_byte, hi_byte per entry, $00 to end
-"""
-
-
-    php
-    sep #0x20  ; 8-bit A
-
-; Build the HDMA data table in WRAM
-; Use UpdateMenuScrollHDMA directly - InitMenuHDMATable is redundant
-; since UpdateMenuScrollHDMA is called immediately after anyway
-    jsr.w update_menu_scroll_hdma
-
-; Configure HDMA channel 5 for DIRECT mode
-; Must use long addressing - DB may be $7E but registers are at $00:43xx
-    lda #0x02  ; Mode: DIRECT, write 2 bytes to same PPU reg
-    sta.l HDMA5_CTRL  ; $004350
-
-    lda #0x0E  ; BG1VOFS register ($210E)
-    sta.l HDMA5_DEST  ; $004351
-
-; Source = HDMA table in WRAM at $7E9800
-    rep #0x20  ; 16-bit A
-    lda.w #MENU_HDMA_TABLE_ADDR  ; $9800
-    sta.l HDMA5_SRC_LO  ; $004352-$004353
-    sep #0x20  ; 8-bit A
-    lda #MENU_HDMA_BANK  ; $7E
-    sta.l HDMA5_SRC_BANK  ; $004354
-
-; HDMA channel 5 is now enabled via shadow variable (menu_hdma_enable)
-; The NMI hook at $8083 reads the shadow and writes to HDMAEN
-
-    plp
-    rts
-
-disable_menu_inventory_hdma:
-"""
-Disables HDMA channel 5 when leaving item menu
-The shadow variable is cleared by menu_exit_hook
-"""
-    php
-    sep #0x20  ; 8-bit A
-    ; Shadow variable cleared by caller, NMI will write 0 to HDMAEN
-    plp
-    rts
-
-init_menu_hdma_table:
-"""
-Builds direct mode HDMA table in WRAM at $7E9800
-Format: count, lo, hi per entry, $00 to end
-Initial state: all rows at BASE scroll (no circular buffer offset)
-"""
-    rep #0x30  ; 16-bit A, X, Y
-    php
-    pha  ; Save A
-    phx  ; Save X
-    phy  ; Save Y
-
-; Save DP bytes we'll use as scratch
-    lda.b 0x40
-    pha  ; Save $40-$41
-
-; X = table write offset
-    ldx.w #0x0000
-
-; Entry 0: Border area - 48 scanlines at BASE scroll
-    sep #0x20  ; 8-bit A for count byte
-    lda #48  ; 48 scanlines
-    sta.l MENU_HDMA_TABLE, x
-    inx
-    rep #0x20  ; 16-bit A for value
-    lda.w menu_rolling_base_scroll
-    sta.l MENU_HDMA_TABLE, x
-    inx
-    inx
-
-; Entries 1-10: Item rows - 16 scanlines each at BASE scroll (initial)
-; For init, all rows use BASE (no circular buffer offset yet)
-    lda.w #0x0000
-    sta.b 0x40  ; Row counter
-
-_init_item_rows:
-    sep #0x20  ; 8-bit A for count
-    lda #16  ; 16 scanlines per item row
-    sta.l MENU_HDMA_TABLE, x
-    inx
-    rep #0x20  ; 16-bit A for value
-    lda.w menu_rolling_base_scroll
-    sta.l MENU_HDMA_TABLE, x
-    inx
-    inx
-
-    inc.b 0x40
-    lda.b 0x40
-
-    cmp.w #1  ; 10 rows
-    bcc _init_item_rows
-
-; Entry 11: Below items - 16 scanlines at BASE + 16
-; Lock to show the bottom window border area
-    sep #0x20
-    lda #16  ; 16 scanlines
-    sta.l MENU_HDMA_TABLE, x
-    inx
-    rep #0x20
-    lda.w menu_rolling_base_scroll
-    clc
-    adc.w #16  ; Lock at base + 16
-    sta.l MENU_HDMA_TABLE, x
-    inx
-    inx
-
-; End marker
-    sep #0x20
-    lda #0x00
-    sta.l MENU_HDMA_TABLE, x
-
-; Restore DP bytes
-    rep #0x20
-    pla
-    sta.b 0x40  ; Restore $40-$41
-
-; Restore registers
-    ply
-    plx
-    pla
-    plp
-    rts
-
-update_menu_scroll_hdma:
-"""Build the field-menu HDMA scroll table via the shared engine."""
-    engine_update_scroll_hdma(menu_rolling, MENU_HDMA_SHADOW, MENU_BUFFER_SLOTS, MENU_VISIBLE_ITEMS, _menu_hdma_header, _menu_hdma_footer, _menu_hdma_signal)  ; noqa: E501
-
-_menu_hdma_header:
-"""Field profile header: 48-scanline border at BASE scroll (one entry)."""
-    sep #0x20
-    lda #48
-    sta.l MENU_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w menu_rolling_base_scroll
-    sta.l MENU_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_menu_hdma_footer:
-"""Field profile footer: 16 scanlines at BASE+16 (locks bottom-border row 24)."""
-    sep #0x20
-    lda #16
-    sta.l MENU_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w menu_rolling_base_scroll
-    clc
-    adc.w #16
-    sta.l MENU_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_menu_hdma_signal:
-"""Field profile NMI signal: set copy-pending shadow flag."""
-    sep #0x20
-    lda #0x01
-    sta.w menu_hdma_copy_pending
-    rts
-
-init_menu_rolling_buffer_impl:
-"""Init field rolling buffer (10 visible, lazy 11th slot)."""
-    engine_init_rolling_buffer(menu_rolling, MENU_VISIBLE_ITEMS, _menu_draw_inventory_window, ensure_hdma_initialized, _menu_render_item_to_slot)  ; noqa: E501
-
-_menu_draw_inventory_window:
-"""Field menu draws the InventoryWindow ($DCCE) frame on entry."""
-    rep #0x10
-    ldy.w #0xDCCE
-    jsr.l draw_window_trampoline
-    sep #0x10
-    rts
-
-menu_scroll_down_prepare:
-"""Field profile scroll-down pre-render."""
-    engine_scroll_down_prepare(menu_rolling, 0x1B1A, MENU_SCROLL_LIMIT, MENU_VISIBLE_ITEMS, MENU_BUFFER_SLOTS, ensure_hdma_initialized, _menu_render_item_to_slot, update_menu_scroll_hdma)  ; noqa: E501
-
-menu_scroll_up_prepare:
-"""Field profile scroll-up pre-render."""
-    engine_scroll_up_prepare(menu_rolling, 0x1B1A, MENU_BUFFER_SLOTS, ensure_hdma_initialized, _menu_render_item_to_slot, update_menu_scroll_hdma)  ; noqa: E501
-
-; _menu_render_item_to_slot
-; Renders an item to a specific circular buffer slot in the tilemap.
-;
-; Input: menu_rolling_edge_row = item index (0-47) for data lookup
-;        menu_rolling_slot_index = slot index (0-11) for destination
-;
-; Strategy: Set up $5d = slot_index (for Y position calculation),
-;           $5a = pointer to item data, then call game's DrawItemSlot.
-
-_menu_render_item_to_slot:
-    php
-    phb
-
-; Set data bank to $7E for WRAM access
-    lda #0x7E
-    pha
-    plb
-
-; Set 16-bit A and X/Y for consistent register handling
-    rep #0x30  ; 16-bit A and X/Y
-
-; Save registers and key direct page variables
-    pha
-    phx  ; Save X (16-bit)
-    phy  ; Save Y (16-bit)
-    lda.b 0x5a
-    pha
-    lda.b 0x29  ; Save tilemap buffer pointer
-    pha
-    lda.b 0x45  ; Save $45-$46 (used by game routines)
-    pha
-    lda.b 0x33  ; Save $33-$34 (used for tile attributes)
-    pha
-    sep #0x20  ; 8-bit A
-    lda.b 0x5d
-    pha
-    lda.b 0xDB  ; Save tile attribute byte
-    pha
-
-; Set $29 = $B600 for BG1 tilemap buffer
-    rep #0x20  ; 16-bit A (X/Y still 16-bit)
-    lda.w #0xB600
-    sta.b 0x29
-    sep #0x20  ; 8-bit A
-
-; Calculate item data pointer: $1440 + (edge_row * Item.__size)
-    lda.w menu_rolling_edge_row
-    asl  ; * Item.__size (2 bytes per Item)
-    clc
-    adc #0x40  ; Low byte of $1440
-    sta.b 0x5a
-    lda #0x14  ; High byte of $1440
-    adc #0x00  ; Add carry
-    sta.b 0x5b
-
-; Load Item.id and Item.qty from ($5A) via long-addressing into WRAM.
-    rep #0x20
-    lda.b 0x5a  ; Pointer value = $1440 + edge_row * Item.__size
-    tax
-    sep #0x20
-    lda.l 0x7E0000 + Item.id, x
-    pha  ; Save Item.id for CheckCanUseItem
-    lda.l 0x7E0000 + Item.qty, x
-    sta.b 0x5C  ; Store qty in $5C
-
-; Call CheckCanUseItem to set palette in $DB
-; Input: A = item ID
-; Output: $DB = $00 (usable) or $04 (not usable)
-    stz.b 0x34  ; Clear $34 (no priority/flip bits)
-    pla  ; Restore item ID
-    jsr.l check_can_use_item_trampoline  ; bank-$01 trampoline for original @ $A25D (sets $DB)
-
-; Set $5d = slot_index (for AND #$01 check, but we patched to AND #$00)
-    lda.w menu_rolling_slot_index
-    sta.b 0x5d
-
-; Calculate Y = slot_index * 128 + 70
-; Y is the tilemap offset for this slot
-; +64 for window border (1 tile row = 32 tiles × 2 bytes)
-; +6 for left margin (3 tiles)
-    rep #0x20  ; 16-bit A (X/Y already 16-bit)
-    lda.w menu_rolling_slot_index
-    and.w #0x00FF  ; Clear high byte
-    xba  ; Swap bytes: A = slot * 256
-    lsr  ; A = slot * 128
-    clc
-    adc.w #0x0044  ; + 68 (64 border + 4 margin)
-    tay  ; Y = tilemap offset (16-bit transfer)
-    sep #0x20  ; 8-bit A
-
-; Check for trash can item ($FF) - needs special 2x2 tile graphic
-    lda (0x5a)  ; Load item ID
-    cmp #0xFF
-    bne _not_trash_item
-    jsr.w draw_trash_single_column  ; Draw trash icon
-    bra _skip_draw_item_slot
-
-_not_trash_item:
-    ; Clear the 2x2 trash can area first (in case we scrolled from trash position)
-    jsr.w _clear_trash_area
-
-; Call DrawItemSlot inner at $A1ED
-; Expects: Y = tilemap offset, ($5A) = item pointer, ($29) = tilemap base
-    jsr.l draw_item_slot_inner_trampoline  ; bank-$01 trampoline for original @ $A1ED
-
-_skip_draw_item_slot:
-
-; Restore direct page variables and registers (reverse order)
-    pla
-    sta.b 0xDB  ; Restore tile attribute byte
-    pla
-    sta.b 0x5d
-    rep #0x20  ; 16-bit A
-    pla
-    sta.b 0x33  ; Restore $33-$34 (tile attributes)
-    pla
-    sta.b 0x45  ; Restore $45-$46 (used by game routines)
-    pla
-    sta.b 0x29  ; Restore tilemap buffer pointer
-    pla
-    sta.b 0x5a
-    rep #0x10  ; 16-bit X/Y for pop (match push)
-    ply  ; Restore Y (16-bit)
-    plx  ; Restore X (16-bit)
-    pla  ; Restore A (16-bit - still in 16-bit A from above)
-
-    plb
-    plp
-    rts
-
-ensure_hdma_initialized:
-"""
-Lazy initialization: captures $93 and sets up HDMA on first scroll.
-Called from scroll prepare functions.
-Checks if base_scroll == 0xFFFF (sentinel) and if so, initializes.
-"""
-
-; Check if already initialized (base_scroll != 0xFFFF)
-    rep #0x20  ; 16-bit A
-    lda.w menu_rolling_base_scroll
-    cmp.w #0xFFFF
-    bne _hdma_already_init
-
-; Capture base scroll from $0193 (BG1VOFS shadow, menu uses DP=$0100)
-; Use long addressing to ensure we read from WRAM
-    .db 0xAF  ; LDA.L opcode
-    .db 0x93, 0x01, 0x7E  ; $7E0193
-    sta.w menu_rolling_base_scroll
-
-; Initialize HDMA channel configuration
-    sep #0x20  ; Back to 8-bit for InitMenuInventoryHDMA
-    jsr.w init_menu_inventory_hdma
-
-; NOW enable HDMA via shadow variable (channel is configured)
-; Force long addressing: STA.L $7E1BAE
-    lda #0x20  ; Channel 5
-    .db 0x8F  ; STA.L opcode
-    .dw menu_hdma_enable  ; $1BAE
-    .db 0x7E  ; Bank $7E
-    rts
-
-_hdma_already_init:
-    sep #0x20  ; Restore 8-bit mode
-    rts
-
-; STATE MACHINE ROUTINES (FF6-style non-blocking scroll)
-
-scroll_state_check:
-"""
-Called at main loop entry ($019FF2) to handle scroll animation frames.
-If scrolling is active, processes one frame and skips input handling.
-
-Returns: Carry clear = process input normally
- Carry set = skip input (still scrolling)
-"""
-
-
-    php
-    sep #0x20  ; 8-bit A
-
-; Check if we're scrolling
-    lda.w menu_scroll_state
-    beq _scroll_state_idle
-
-; We're scrolling - process one animation frame
-    jsr.l update_scroll_frame_impl
-
-; Check if scroll finished
-    lda.w menu_scroll_remaining
-    bne _scroll_still_active
-
-; Scroll finished - clean up and return to idle
-    jsr.l finish_scroll_impl
-
-_scroll_state_idle:
-    plp
-    clc  ; Carry clear = process input
-    rts
-
-_scroll_still_active:
-    plp
-    sec  ; Carry set = skip input
-    rts
-
-start_scroll_down_impl:
-"""Field profile: kick scroll-down state machine."""
-    engine_start_scroll_down(menu_rolling, 0x1B1A, MENU_VISIBLE_ITEMS, MENU_BUFFER_SLOTS, SCROLL_TOTAL_PIXELS, SCROLL_PIXELS_PER_FRAME, ensure_hdma_initialized, _menu_render_item_to_slot, update_menu_scroll_hdma)  ; noqa: E501
-
-start_scroll_up_impl:
-"""Field profile: kick scroll-up state machine."""
-    engine_start_scroll_up(menu_rolling, 0x1B1A, MENU_BUFFER_SLOTS, SCROLL_TOTAL_PIXELS, ensure_hdma_initialized, _menu_render_item_to_slot, update_menu_scroll_hdma)  ; noqa: E501
-
-update_scroll_frame_impl:
-"""Field profile: per-frame scroll animation tick."""
-    engine_update_scroll_frame(menu_rolling, SCROLL_PIXELS_PER_FRAME, update_menu_scroll_hdma)
-
-finish_scroll_impl:
-"""Field profile: end-of-animation pre-render + cleanup."""
-    engine_finish_scroll(menu_rolling, 0x1B1A, MENU_VISIBLE_ITEMS, MENU_BUFFER_SLOTS, MENU_TOTAL_ITEMS, _menu_render_item_to_slot, update_menu_scroll_hdma)  ; noqa: E501
-
-; Inventory Rolling Buffer Patches - Relocated to Bank $20
-; These routines are called via JSL from bank $01 hooks.
-
-.if INVENTORY_ROLLING_BUFFER {
-menu_entry_hook_impl:
-"""Field-menu entry hook: lazy-init HDMA + force shadow flush before first frame."""
-    stz.w 0x1B1F
-    lda #0x00
-    sta.l 0x7E1BAE
-; menu_hdma_enable
-    stz.w menu_scroll_state
-    stz.w menu_scroll_remaining
-    stz.w menu_scroll_direction
-    stz.w menu_transfer_pending
-    stz.w menu_hdma_copy_pending
-; Clear HDMA copy flag
-    stz.w menu_scroll_anim_offset
-; Clear low byte
-    stz.w menu_scroll_anim_offset + 1
-; Clear high byte
-; Initialize cursor column to 0 for single-column mode
-; This ensures $1b22 is always 0 even if it had a value from previous menu
-    stz.w 0x1B22
-; cursor_x = 0
-; InitMenuRollingBuffer_Impl is called later via patched JSR at $9F7B
-    rtl
-
-menu_exit_hook_impl:
-"""Field menu teardown: zero HDMA shadow, full 12-byte rolling state, ch5 registers."""
-    php
-    sep #0x20
-    lda #0x00
-    sta.l 0x7E1BAE
-; menu_hdma_enable shadow off so NMI writes 0 to HDMAEN this frame
-    sta.l 0x004350
-    sta.l 0x004351
-    sta.l 0x004352
-    sta.l 0x004353
-    sta.l 0x004354
-; HDMA5 ctrl/dest/src cleared so a stale config can't restart on next mode switch
-    rep #0x20
-    lda.w #0x0000
-    sta.w menu_rolling
-    sta.w menu_rolling + 2
-    sta.w menu_rolling + 4
-    sta.w menu_rolling + 6
-    sta.w menu_rolling + 8
-    sta.w menu_rolling + 10
-    plp
-    jsr.l reset_sprites_trampoline
-    rtl
-; swap_redraw_hook_impl_body
-; Called after item swap to redraw visible items correctly.
-; Must render to the correct circular buffer slots based on current buffer_pos.
-; Does NOT reset buffer_pos - we stay at the current scroll position.
-
-swap_redraw_hook_impl_body:
-"""Field profile: post-swap re-render of all 11 slots."""
-    engine_swap_redraw(menu_rolling, 0x1B1A, MENU_BUFFER_SLOTS, MENU_TOTAL_ITEMS, ensure_hdma_initialized, _menu_render_item_to_slot, _clear_inventory_slot, update_menu_scroll_hdma)  ; noqa: E501
-
-
-; _clear_inventory_slot
-; Clears a single inventory slot in the tilemap buffer.
-; Input: menu_rolling_slot_index = slot to clear (0-10)
-; Used when item index is out of bounds (>= 48)
-
-_clear_inventory_slot:
-    php
-    phb
-; Set data bank to $7E for WRAM access
-    lda #0x7E
-    pha
-    plb
-; Save registers - use 16-bit mode for consistent push/pop
-    rep #0x30
-; 16-bit A and X/Y
-    pha
-    phx
-    phy
-    lda.b 0x29
-    pha
-; Set $29 = $B600 for BG1 tilemap buffer
-    lda.w #0xB600
-    sta.b 0x29
-; Calculate Y = slot_index * 128 + 70
-    lda.w menu_rolling_slot_index
-    and.w #0x00FF
-    xba
-; A = slot * 256
-    lsr
-; A = slot * 128
-    clc
-    adc.w #0x0044
-; + 68 (64 border + 4 margin)
-    tay
-; Y = tilemap offset (16-bit)
-    sep #0x20
-; 8-bit A for tile writes (X/Y stay 16-bit)
-; Clear the item name area (12 tiles = 24 bytes)
-; Use $FF as blank tile
-; Note: X is 16-bit but we only use low byte; loop works correctly
-    ldx.w #12
-; 12 tiles for item name + quantity (force 16-bit immediate)
-
-_clear_slot_loop:
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-    iny
-    lda #0x04
-; Palette 4 (matches normal items)
-    sta (0x29), y
-    iny
-    dex
-    bne _clear_slot_loop
-; Restore registers - must match push mode
-    rep #0x20
-; 16-bit A for pop (X/Y already 16-bit)
-    pla
-    sta.b 0x29
-    ply
-    plx
-    pla
-    plb
-    plp
-    rts
-; _clear_trash_area
-; Clears the 2x2 trash can area with blank tiles ($FF).
-; Called before drawing normal items to remove any leftover trash icon.
-; Input: Y = tilemap offset
-;        ($29) = tilemap base
-;
-
-_clear_trash_area:
-    phy
-; Save Y
-; First row: 2 tiles
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-; Second row: Y + 64 from start
-    ply
-; Restore original Y
-    phy
-; Save again
-    rep #0x20
-    tya
-    clc
-    adc.w #64
-    tay
-    sep #0x20
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-    ply
-; Restore Y
-    rts
-
-draw_trash_single_column:
-
-
-"""
-Draws the trash can 2x2 tile graphic for single-column inventory.
-Input: Y = tilemap offset (from slot calculation)
-($29) = tilemap base ($B600)
-menu_rolling_slot_index = current slot
-Tiles: $04 (top-left), $05 (top-right), $06 (bottom-left), $07 (bottom-right)
-
-Tilemap format: [tile_number, attributes] pairs
-Each row is 64 bytes (32 tiles × 2 bytes)
-"""
-
-
-; Y points to start of item slot area
-; Draw 2x2 trash can icon, then clear remaining 10 tiles per row
-; Save starting Y for second row calculation
-    rep #0x20
-; 16-bit for push
-    phy
-; Save starting Y
-    sep #0x20
-; 8-bit A
-; First row: tiles $04, $05
-    lda #0x04
-; Tile $04 (top-left of trash can)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute: palette 0, no flip
-    sta (0x29), y
-    iny
-    lda #0x05
-; Tile $05 (top-right)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-; Clear remaining 13 tiles on first row (10 name + colon + 2 digits)
-    ldx.w #13
-
-_clear_row1:
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-    dex
-    bne _clear_row1
-; Restore starting Y and add 64 for second row
-    rep #0x20
-; 16-bit A
-    pla
-; Get starting Y
-    clc
-    adc.w #64
-; +64 bytes = next tilemap row
-    tay
-    sep #0x20
-; 8-bit A
-; Second row: tiles $06, $07
-    lda #0x06
-; Tile $06 (bottom-left)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-    lda #0x07
-; Tile $07 (bottom-right)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-; Clear remaining 13 tiles on second row (10 name + colon + 2 digits)
-    ldx.w #13
-
-_clear_row2:
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-    dex
-    bne _clear_row2
-    rts
-; NmiDmaTransferCheck_Impl moved to bank $20 (battle/inventory_rolling.s)
-; as field_menu_nmi_dma_transfer_check_impl to save space in bank $01
-; check_and_clear_count_impl
-; Called from DrawItemSlot to check item ID and handle count display.
-; - If item ID is 0: writes $FF tiles to clear count, skips to RTS
-; - If item ID is $FE: skips to RTS (no clearing needed)
-; - Otherwise: returns normally to draw count
-;
-; Input: $5a = pointer to item data, Y = tilemap offset, $29 = tilemap ptr
-; Modifies: A, return address on stack if skipping
-
-check_and_clear_count_impl:
-"""Drop the count column for empty/used slots."""
-    lda (0x5a)
-; Load item ID
-    beq _clear_count
-; If 0, clear and skip
-    cmp #0xFE
-; Check for special item $FE
-    bne _normal_return
-; If not $FE, return normally to draw count
-; Item is $FE - skip count but don't clear
-    bra _skip_to_rts
-
-_clear_count:
-; Write $FF (blank tiles) to count area: colon + 2 digits = 3 tiles
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-; Colon position
-    iny
-    lda.b 0xdb
-; Attribute byte
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-; First digit position
-    iny
-    lda.b 0xdb
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-; Second digit position
-    iny
-    lda.b 0xdb
-    sta (0x29), y
-
-_skip_to_rts:
-; Replace the entire DrawItemSlot return chain so we land on the RTS at $01:A222
-; instead of falling back to $A205 (NOPs + colon write).
-;
-; Stack on entry to Impl (pushed by trampoline + caller, top first):
-;   [JSL Impl return: PCL, PCH, PB ($01)]
-;   [JSR check_and_clear_count return: PCL, PCH ($A204)]
-;   [DrawItemSlot caller's return ...]
-;
-; Pop both the JSL Impl return (3 bytes) AND the JSR trampoline return (2 bytes),
-; then push a synthetic return to $01:A222 (RTS) for our RTL.
-    pla  ; PCL of JSL Impl return
-    pla  ; PCH of JSL Impl return
-    pla  ; PB  of JSL Impl return
-    pla  ; PCL of JSR check_and_clear_count return
-    pla  ; PCH of JSR check_and_clear_count return
-; Stack top now = DrawItemSlot's own caller return.
-; Push return for RTL: $01:A221 (so RTL lands PC=$01:A222 = RTS).
-    lda #0x01
-    pha
-    lda #0xA2
-    pha
-    lda #0x21
-    pha
-
-_normal_return:
-    rtl
-
-circular_slot_calc:
-
-
-"""
-Calculate tilemap Y offset using circular buffer position.
-Called from patched code at $A1BA via CircularSlotCalc_ext.
-
-Input: $5D = game's slot counter (0, 2, 4, 6... incremented by 2 per row)
-$5A = src pointer ($1440 = inventory, $FF28 = treasure drops)
-Output: Y = tilemap offset for circular buffer slot
-Preserves: 16-bit A mode on exit
-"""
-
-
-    sep #0x20
-; 8-bit A
-; Drops list at $FF28 reuses _a181 but doesn't have any rolling state.
-; The original 2-col Y calc collides slots after we forced col=0 globally
-; via the `and #$00` patch at DrawItemSlot, so two drops would render at
-; the same scanline. Detect drops via $5B high byte and emit a unique
-; per-slot Y (slot * 64 + 4) instead of any circular math.
-    lda.b 0x5b
-    cmp #0xFF
-    bne _circ_slot_not_drops
-    lda.b 0x5d
-    rep #0x20
-    and.w #0x00FF
-    asl
-    asl
-    asl
-    asl
-    asl
-    asl  ; * 64. $5D increments by 2 → step = 128 bytes = 2 tilemap rows,
-; matching the 16-px-per-item layout original uses for col-0
-; drops with BG1VOFS=-32 ($93) pushing them down to scanline 32.
-    clc
-    adc.w #0x0044  ; col 2 + 1 tilemap row down (+$40) so the first slot
-; clears the window header before BG1VOFS shifts it.
-    tay
-    rts
-_circ_slot_not_drops:
-; Inventory in treasure context uses treasure_rolling_buffer_pos.
-    lda.l 0x7E0000 + 0x1BD6  ; treasure_hdma_enable
-    beq _circ_check_field
-    lda.b 0x5d
-    lsr
-    clc
-    adc.l 0x7E0000 + 0x1BD1  ; treasure_rolling_buffer_pos
-_t_circ_mod:
-    cmp #6  ; TREASURE_BUFFER_SLOTS
-    bcc _t_circ_done
-    sec
-    sbc #6
-    bra _t_circ_mod
-_t_circ_done:
-    rep #0x20
-    and.w #0x00FF
-    xba
-    lsr
-    clc
-    adc.w #0x0004
-    tay
-    rts
-_circ_check_field:
-; Check if circular buffer mode is active (HDMA enabled)
-    lda.l 0x7E0000 + menu_hdma_enable
-    beq _circ_slot_original
-; Not active, use original calculation
-; Circular buffer Y calculation
-; NOTE: Game increments $5D by 2 for each row (0, 2, 4, 6, 8, 10, 12, 14, 16, 18)
-; We must divide by 2 first to get the visual slot (0-9)
-    lda.b 0x5d
-; Load slot counter (0, 2, 4...)
-    lsr
-; Divide by 2 to get visual slot (0-9)
-    clc
-    adc.l 0x7E0000 + menu_rolling_buffer_pos
-; Add buffer_pos
-
-_circ_slot_mod:
-    cmp #MENU_BUFFER_SLOTS
-; >= 11?
-    bcc _circ_slot_done
-    sec
-    sbc #MENU_BUFFER_SLOTS
-; Subtract 11 to wrap
-    bra _circ_slot_mod
-
-_circ_slot_done:
-; A = circular slot (0-10)
-    rep #0x20
-; 16-bit A
-    and.w #0x00FF
-; Clear high byte (force 16-bit immediate)
-    xba
-; A = slot * 256
-    lsr
-; A = slot * 128
-    clc
-    adc.w #0x0004
-; + 4 (margin only, $29 already includes border)
-    tay
-; Y = tilemap offset
-    rts
-
-_circ_slot_original:
-; Original game calculation: Y = ($5D / 2) * 128 + 4
-    lda.b 0x5d
-    lsr
-; /2
-    rep #0x20
-; 16-bit A
-    and.w #0x00FF
-; Clear high byte (force 16-bit immediate)
-    xba
-; *256
-    lsr
-; *128
-    clc
-    adc.w #0x0004
-; +4
-    tay
-    rts
-
-circular_slot_calc_ext:
-"""Trampoline to call CircularSlotCalc from bank $01 patch at $A1BA"""
-    jsr.w circular_slot_calc
-    rtl
+.include "../bank20.i"
+
+.alloc inventory_rolling_block in bank20_reloc {
+    init_menu_inventory_hdma:
+    """
+    Sets up HDMA channel 5 for per-scanline BG1 vertical scroll control
+    Called when entering the item menu
+
+    HDMA mode: $02 = write 2 bytes to same register, DIRECT mode (like FF6)
+    Register: $210E (BG1VOFS)
+    Table format: count_byte, lo_byte, hi_byte per entry, $00 to end
+    """
+
+
+        php
+        sep #0x20  ; 8-bit A
+
+    ; Build the HDMA data table in WRAM
+    ; Use UpdateMenuScrollHDMA directly - InitMenuHDMATable is redundant
+    ; since UpdateMenuScrollHDMA is called immediately after anyway
+        jsr.w update_menu_scroll_hdma
+
+    ; Configure HDMA channel 5 for DIRECT mode
+    ; Must use long addressing - DB may be $7E but registers are at $00:43xx
+        lda #0x02  ; Mode: DIRECT, write 2 bytes to same PPU reg
+        sta.l HDMA5_CTRL  ; $004350
+
+        lda #0x0E  ; BG1VOFS register ($210E)
+        sta.l HDMA5_DEST  ; $004351
+
+    ; Source = HDMA table in WRAM at $7E9800
+        rep #0x20  ; 16-bit A
+        lda.w #MENU_HDMA_TABLE_ADDR  ; $9800
+        sta.l HDMA5_SRC_LO  ; $004352-$004353
+        sep #0x20  ; 8-bit A
+        lda #MENU_HDMA_BANK  ; $7E
+        sta.l HDMA5_SRC_BANK  ; $004354
+
+    ; HDMA channel 5 is now enabled via shadow variable (menu_hdma_enable)
+    ; The NMI hook at $8083 reads the shadow and writes to HDMAEN
+
+        plp
+        rts
+
+    disable_menu_inventory_hdma:
+    """
+    Disables HDMA channel 5 when leaving item menu
+    The shadow variable is cleared by menu_exit_hook
+    """
+        php
+        sep #0x20  ; 8-bit A
+        ; Shadow variable cleared by caller, NMI will write 0 to HDMAEN
+        plp
+        rts
+
+    init_menu_hdma_table:
+    """
+    Builds direct mode HDMA table in WRAM at $7E9800
+    Format: count, lo, hi per entry, $00 to end
+    Initial state: all rows at BASE scroll (no circular buffer offset)
+    """
+        rep #0x30  ; 16-bit A, X, Y
+        php
+        pha  ; Save A
+        phx  ; Save X
+        phy  ; Save Y
+
+    ; Save DP bytes we'll use as scratch
+        lda.b 0x40
+        pha  ; Save $40-$41
+
+    ; X = table write offset
+        ldx.w #0x0000
+
+    ; Entry 0: Border area - 48 scanlines at BASE scroll
+        sep #0x20  ; 8-bit A for count byte
+        lda #48  ; 48 scanlines
+        sta.l MENU_HDMA_TABLE, x
+        inx
+        rep #0x20  ; 16-bit A for value
+        lda.w menu_rolling_base_scroll
+        sta.l MENU_HDMA_TABLE, x
+        inx
+        inx
+
+    ; Entries 1-10: Item rows - 16 scanlines each at BASE scroll (initial)
+    ; For init, all rows use BASE (no circular buffer offset yet)
+        lda.w #0x0000
+        sta.b 0x40  ; Row counter
+
+    _init_item_rows:
+        sep #0x20  ; 8-bit A for count
+        lda #16  ; 16 scanlines per item row
+        sta.l MENU_HDMA_TABLE, x
+        inx
+        rep #0x20  ; 16-bit A for value
+        lda.w menu_rolling_base_scroll
+        sta.l MENU_HDMA_TABLE, x
+        inx
+        inx
+
+        inc.b 0x40
+        lda.b 0x40
+
+        cmp.w #1  ; 10 rows
+        bcc _init_item_rows
+
+    ; Entry 11: Below items - 16 scanlines at BASE + 16
+    ; Lock to show the bottom window border area
+        sep #0x20
+        lda #16  ; 16 scanlines
+        sta.l MENU_HDMA_TABLE, x
+        inx
+        rep #0x20
+        lda.w menu_rolling_base_scroll
+        clc
+        adc.w #16  ; Lock at base + 16
+        sta.l MENU_HDMA_TABLE, x
+        inx
+        inx
+
+    ; End marker
+        sep #0x20
+        lda #0x00
+        sta.l MENU_HDMA_TABLE, x
+
+    ; Restore DP bytes
+        rep #0x20
+        pla
+        sta.b 0x40  ; Restore $40-$41
+
+    ; Restore registers
+        ply
+        plx
+        pla
+        plp
+        rts
+
+    update_menu_scroll_hdma:
+    """Build the field-menu HDMA scroll table via the shared engine."""
+        ; Build field-menu HDMA scroll table. Inlined from the former
+        ; engine_update_scroll_hdma macro for the same reason as treasure.
+    {
+        php
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x40
+        pha
+        lda.b 0x42
+        pha
+        ldx.w #0x0000
+        jsr.w _menu_hdma_header
+        stz.b 0x42
+
+    _row_loop:
+        lda.w menu_rolling + RollingBufferState.buffer_pos
+        and.w #0x00FF
+        clc
+        adc.b 0x42
+
+    _mod_loop:
+        cmp.w #MENU_BUFFER_SLOTS
+        bcc _mod_done
+        sec
+        sbc.w #MENU_BUFFER_SLOTS
+        bra _mod_loop
+
+    _mod_done:
+        asl
+        asl
+        asl
+        asl
+        sta.b 0x40
+        lda.b 0x42
+        and.w #0x00FF
+        asl
+        asl
+        asl
+        asl
+        eor.w #0xFFFF
+        inc
+        clc
+        adc.b 0x40
+        clc
+        adc.w menu_rolling + RollingBufferState.base_scroll
+        sta.b 0x40
+        sep #0x20
+        lda #16
+        sta.l MENU_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.b 0x40
+        sta.l MENU_HDMA_SHADOW, x
+        inx
+        inx
+        rep #0x20
+        inc.b 0x42
+        lda.b 0x42
+        cmp.w #MENU_VISIBLE_ITEMS
+        bcs _row_loop_done
+        jmp.w _row_loop
+
+    _row_loop_done:
+        jsr.w _menu_hdma_footer
+        sep #0x20
+        lda #0x00
+        sta.l MENU_HDMA_SHADOW, x
+        jsr.w _menu_hdma_signal
+        rep #0x20
+        pla
+        sta.b 0x42
+        pla
+        sta.b 0x40
+        ply
+        plx
+        pla
+        plp
+        rts
+    }
+
+    _menu_hdma_header:
+    """Field profile header: 48-scanline border at BASE scroll (one entry)."""
+        sep #0x20
+        lda #48
+        sta.l MENU_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w menu_rolling_base_scroll
+        sta.l MENU_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _menu_hdma_footer:
+    """Field profile footer: 16 scanlines at BASE+16 (locks bottom-border row 24)."""
+        sep #0x20
+        lda #16
+        sta.l MENU_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w menu_rolling_base_scroll
+        clc
+        adc.w #16
+        sta.l MENU_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _menu_hdma_signal:
+    """Field profile NMI signal: set copy-pending shadow flag."""
+        sep #0x20
+        lda #0x01
+        sta.w menu_hdma_copy_pending
+        rts
+
+    init_menu_rolling_buffer_impl:
+    """
+    Init field rolling buffer (10 visible, lazy 11th slot).
+
+    Populates the phase-1 engine config + hook fields so the bank-20
+    `rolling_engine_*` entries can drive the same instance the macro
+    expansion below already manages. Engine consumers stay inert until
+    phase 2 swaps the macro for the JSL path  ; populating now lets the
+    integration tests assert real config values and gives us a stable
+    ABI surface to port against.
+    """
+
+
+        php
+        rep #0x30  ; M=16, X=16
+        sep #0x20  ; M=8 (X stays 16)
+        ; visible_rows + slot_height_tiles
+        lda.b #MENU_VISIBLE_ITEMS
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.visible_rows
+        lda.b #0x02  ; 2 BG rows per slot
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.slot_height_tiles
+        ; item_list_ptr = $7E:1440 (vanilla inventory array)
+        lda.b #0x40
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.item_list_ptr
+        lda.b #0x14
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.item_list_ptr + 1
+        lda.b #0x7E
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.item_list_ptr + 2
+        ; item_count = 48 (vanilla field inventory)
+        lda.b #0x30
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.item_count
+        ; hdma_channel = 5 (BG1VOFS HDMA used by field-items rolling buffer)
+        lda.b #0x05
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.hdma_channel
+        ; vwf_cfg_ptr = $70:7080 (VWF_CONFIG_BASE)
+        lda.b #0x80
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.vwf_cfg_ptr
+        lda.b #0x70
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.vwf_cfg_ptr + 1
+        lda.b #0x70
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.vwf_cfg_ptr + 2
+        ; fn_render_slot = menu_fn_render_slot_trampoline (bank-20 RTL wrapper)
+        lda.b #menu_fn_render_slot_trampoline & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_render_slot
+        lda.b #( menu_fn_render_slot_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_render_slot + 1
+        lda.b #( menu_fn_render_slot_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_render_slot + 2
+        ; fn_update_hdma = menu_fn_update_hdma_trampoline
+        lda.b #menu_fn_update_hdma_trampoline & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_update_hdma
+        lda.b #( menu_fn_update_hdma_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_update_hdma + 1
+        lda.b #( menu_fn_update_hdma_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_update_hdma + 2
+        ; fn_draw_window = menu_fn_draw_window_trampoline
+        lda.b #menu_fn_draw_window_trampoline & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_draw_window
+        lda.b #( menu_fn_draw_window_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_draw_window + 1
+        lda.b #( menu_fn_draw_window_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.fn_draw_window + 2
+        lda.b #ROLLING_MENU_ID_FIELD
+        sta.l 0x7E0000 + menu_rolling + RollingBufferState.menu_id
+        plp
+        php
+        rep #0x10
+        ldx.w #menu_rolling
+        jsr.l rolling_engine.rolling_engine_init
+        plp
+        rtl
+
+    menu_fn_render_slot_trampoline:
+    """
+    Bank-20 RTL wrapper around `_menu_render_item_to_slot`.
+
+    php/plp keeps M/X flag mutations inside the inner routine from
+    leaking back to the engine. The engine relies on X-flag = 16
+    throughout init  ; hooks that toggle X-flag truncate X to its low
+    byte, which sends subsequent `sta.l abs,X` writes to garbage
+    addresses.
+    """
+
+
+        php
+        jsr.w _menu_render_item_to_slot
+        plp
+        rtl
+
+    menu_fn_update_hdma_trampoline:
+    """
+    Bank-20 RTL wrapper around `ensure_hdma_initialized`.
+
+    Named fn_update_hdma in the struct but semantically the
+    "arm-HDMA-channel + first-frame setup" hook for init. Phase 3 may
+    split into ensure-once vs per-scroll-update if scroll-tick needs a
+    distinct entry, but the existing field-menu macro chain treats
+    ensure as idempotent so reusing it for both is fine.
+
+    php/plp guards the engine's X-flag = 16 from inner sep/rep.
+    """
+
+
+        php
+        jsr.w ensure_hdma_initialized
+        plp
+        rtl
+
+    menu_fn_draw_window_trampoline:
+    """
+    Bank-20 RTL wrapper around `_menu_draw_inventory_window`.
+
+    php/plp protects the engine's X-flag = 16  ; the inner draw does
+    `rep #$10  ; jsr ; sep #$10` for its own 8-bit-X dance and would
+    otherwise leave X-flag = 8 when we return, corrupting `,X` indexed
+    state writes downstream of the hook in the engine_init body.
+    """
+
+
+        php
+        jsr.w _menu_draw_inventory_window
+        plp
+        rtl
+
+    _menu_draw_inventory_window:
+    """Field menu draws the InventoryWindow ($DCCE) frame on entry."""
+        rep #0x10
+        ldy.w #0xDCCE
+        jsr.l draw_window_trampoline
+        sep #0x10
+        rts
+
+    ; _menu_render_item_to_slot
+    ; Renders an item to a specific circular buffer slot in the tilemap.
+    ;
+    ; Input: menu_rolling_edge_row = item index (0-47) for data lookup
+    ;        menu_rolling_slot_index = slot index (0-11) for destination
+    ;
+    ; Strategy: Set up $5d = slot_index (for Y position calculation),
+    ;           $5a = pointer to item data, then call game's DrawItemSlot.
+
+    _menu_render_item_to_slot:
+        php
+        phb
+
+    ; Set data bank to $7E for WRAM access
+        lda #0x7E
+        pha
+        plb
+
+    ; Set 16-bit A and X/Y for consistent register handling
+        rep #0x30  ; 16-bit A and X/Y
+
+    ; Save registers and key direct page variables
+        pha
+        phx  ; Save X (16-bit)
+        phy  ; Save Y (16-bit)
+        lda.b 0x5a
+        pha
+        lda.b 0x29  ; Save tilemap buffer pointer
+        pha
+        lda.b 0x45  ; Save $45-$46 (used by game routines)
+        pha
+        lda.b 0x33  ; Save $33-$34 (used for tile attributes)
+        pha
+        sep #0x20  ; 8-bit A
+        lda.b 0x5d
+        pha
+        lda.b 0xDB  ; Save tile attribute byte
+        pha
+
+    ; Set $29 = $B600 for BG1 tilemap buffer
+        rep #0x20  ; 16-bit A (X/Y still 16-bit)
+        lda.w #0xB600
+        sta.b 0x29
+        sep #0x20  ; 8-bit A
+
+    ; Calculate item data pointer: $1440 + (edge_row * Item.__size)
+        lda.w menu_rolling_edge_row
+        asl  ; * Item.__size (2 bytes per Item)
+        clc
+        adc #0x40  ; Low byte of $1440
+        sta.b 0x5a
+        lda #0x14  ; High byte of $1440
+        adc #0x00  ; Add carry
+        sta.b 0x5b
+
+    ; Load Item.id and Item.qty from ($5A) via long-addressing into WRAM.
+        rep #0x20
+        lda.b 0x5a  ; Pointer value = $1440 + edge_row * Item.__size
+        tax
+        sep #0x20
+        lda.l 0x7E0000 + Item.id, x
+        pha  ; Save Item.id for CheckCanUseItem
+        lda.l 0x7E0000 + Item.qty, x
+        sta.b 0x5C  ; Store qty in $5C
+
+    ; Call CheckCanUseItem to set palette in $DB
+    ; Input: A = item ID
+    ; Output: $DB = $00 (usable) or $04 (not usable)
+        stz.b 0x34  ; Clear $34 (no priority/flip bits)
+        pla  ; Restore item ID
+        jsr.l check_can_use_item_trampoline  ; bank-$01 trampoline for original @ $A25D (sets $DB)
+
+    ; Set $5d = slot_index (for AND #$01 check, but we patched to AND #$00)
+        lda.w menu_rolling_slot_index
+        sta.b 0x5d
+
+    ; Calculate Y = slot_index * 128 + 70
+    ; Y is the tilemap offset for this slot
+    ; +64 for window border (1 tile row = 32 tiles × 2 bytes)
+    ; +6 for left margin (3 tiles)
+        rep #0x20  ; 16-bit A (X/Y already 16-bit)
+        lda.w menu_rolling_slot_index
+        and.w #0x00FF  ; Clear high byte
+        xba  ; Swap bytes: A = slot * 256
+        lsr  ; A = slot * 128
+        clc
+        adc.w #0x0044  ; + 68 (64 border + 4 margin)
+        tay  ; Y = tilemap offset (16-bit transfer)
+        sep #0x20  ; 8-bit A
+
+    ; Check for trash can item ($FF) - needs special 2x2 tile graphic
+        lda (0x5a)  ; Load item ID
+        cmp #0xFF
+        bne _not_trash_item
+        jsr.w draw_trash_single_column  ; Draw trash icon
+        bra _skip_draw_item_slot
+
+    _not_trash_item:
+        ; Clear the 2x2 trash can area first (in case we scrolled from trash position)
+        jsr.w _clear_trash_area
+
+    ; Call DrawItemSlot inner at $A1ED
+    ; Expects: Y = tilemap offset, ($5A) = item pointer, ($29) = tilemap base
+        jsr.l draw_item_slot_inner_trampoline  ; bank-$01 trampoline for original @ $A1ED
+
+    _skip_draw_item_slot:
+
+    ; Restore direct page variables and registers (reverse order)
+        pla
+        sta.b 0xDB  ; Restore tile attribute byte
+        pla
+        sta.b 0x5d
+        rep #0x20  ; 16-bit A
+        pla
+        sta.b 0x33  ; Restore $33-$34 (tile attributes)
+        pla
+        sta.b 0x45  ; Restore $45-$46 (used by game routines)
+        pla
+        sta.b 0x29  ; Restore tilemap buffer pointer
+        pla
+        sta.b 0x5a
+        rep #0x10  ; 16-bit X/Y for pop (match push)
+        ply  ; Restore Y (16-bit)
+        plx  ; Restore X (16-bit)
+        pla  ; Restore A (16-bit - still in 16-bit A from above)
+
+        plb
+        plp
+        rts
+
+    ensure_hdma_initialized:
+    """
+    Lazy initialization: captures $93 and sets up HDMA on first scroll.
+    Called from scroll prepare functions.
+    Checks if base_scroll == 0xFFFF (sentinel) and if so, initializes.
+    """
+
+    ; Check if already initialized (base_scroll != 0xFFFF)
+        rep #0x20  ; 16-bit A
+        lda.w menu_rolling_base_scroll
+        cmp.w #0xFFFF
+        bne _hdma_already_init
+
+    ; Capture base scroll from $0193 (BG1VOFS shadow, menu uses DP=$0100)
+    ; Use long addressing to ensure we read from WRAM
+        .db 0xAF  ; LDA.L opcode
+        .db 0x93, 0x01, 0x7E  ; $7E0193
+        sta.w menu_rolling_base_scroll
+
+    ; Initialize HDMA channel configuration
+        sep #0x20  ; Back to 8-bit for InitMenuInventoryHDMA
+        jsr.w init_menu_inventory_hdma
+
+    ; NOW enable HDMA via shadow variable (channel is configured)
+    ; Force long addressing: STA.L $7E1BAE
+        lda #0x20  ; Channel 5
+        .db 0x8F  ; STA.L opcode
+        .dw menu_hdma_enable  ; $1BAE
+        .db 0x7E  ; Bank $7E
+        rts
+
+    _hdma_already_init:
+        sep #0x20  ; Restore 8-bit mode
+        rts
+
+    ; STATE MACHINE ROUTINES (FF6-style non-blocking scroll)
+
+    scroll_state_check:
+    """
+    Called at main loop entry ($019FF2) to handle scroll animation frames.
+    If scrolling is active, processes one frame and skips input handling.
+
+    Returns: Carry clear = process input normally
+     Carry set = skip input (still scrolling)
+    """
+
+
+        php
+        sep #0x20  ; 8-bit A
+
+    ; Check if we're scrolling
+        lda.w menu_scroll_state
+        beq _scroll_state_idle
+
+    ; We're scrolling - process one animation frame
+        jsr.l update_scroll_frame_impl
+
+    ; Check if scroll finished
+        lda.w menu_scroll_remaining
+        bne _scroll_still_active
+
+    ; Scroll finished - clean up and return to idle
+        jsr.l finish_scroll_impl
+
+    _scroll_state_idle:
+        plp
+        clc  ; Carry clear = process input
+        rts
+
+    _scroll_still_active:
+        plp
+        sec  ; Carry set = skip input
+        rts
+
+    start_scroll_down_impl:
+    """Field profile: kick scroll-down state machine via the bank-20 engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1B1A  ; field scroll_pos
+        ldx.w #menu_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_down
+        plp
+        rtl
+
+    start_scroll_up_impl:
+    """Field profile: kick scroll-up state machine via the bank-20 engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1B1A
+        ldx.w #menu_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_up
+        plp
+        rtl
+
+    update_scroll_frame_impl:
+    """Field profile: per-frame scroll animation tick via the bank-20 engine."""
+        php
+        rep #0x10
+        ldx.w #menu_rolling
+        jsr.l rolling_engine.rolling_engine_update_scroll_frame
+        plp
+        rtl
+
+    finish_scroll_impl:
+    """Field profile: end-of-animation cleanup via the bank-20 engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1B1A
+        ldx.w #menu_rolling
+        jsr.l rolling_engine.rolling_engine_finish_scroll
+        plp
+        rtl
+
+    ; Inventory Rolling Buffer Patches - Relocated to Bank $20
+    ; These routines are called via JSL from bank $01 hooks.
+
+    .if INVENTORY_ROLLING_BUFFER {
+    menu_entry_hook_impl:
+    """Field-menu entry hook: lazy-init HDMA + force shadow flush before first frame."""
+        stz.w 0x1B1F
+        lda #0x00
+        sta.l 0x7E1BAE
+    ; menu_hdma_enable
+        stz.w menu_scroll_state
+        stz.w menu_scroll_remaining
+        stz.w menu_scroll_direction
+        stz.w menu_transfer_pending
+        stz.w menu_hdma_copy_pending
+    ; Clear HDMA copy flag
+        stz.w menu_scroll_anim_offset
+    ; Clear low byte
+        stz.w menu_scroll_anim_offset + 1
+    ; Clear high byte
+    ; Initialize cursor column to 0 for single-column mode
+    ; This ensures $1b22 is always 0 even if it had a value from previous menu
+        stz.w 0x1B22
+    ; cursor_x = 0
+    ; InitMenuRollingBuffer_Impl is called later via patched JSR at $9F7B
+        rtl
+
+    menu_exit_hook_impl:
+    """Field menu teardown: zero HDMA shadow, full 12-byte rolling state, ch5 registers."""
+        php
+        sep #0x20
+        lda #0x00
+        sta.l 0x7E1BAE
+    ; menu_hdma_enable shadow off so NMI writes 0 to HDMAEN this frame
+        sta.l 0x004350
+        sta.l 0x004351
+        sta.l 0x004352
+        sta.l 0x004353
+        sta.l 0x004354
+    ; HDMA5 ctrl/dest/src cleared so a stale config can't restart on next mode switch
+        rep #0x20
+        lda.w #0x0000
+        sta.w menu_rolling
+        sta.w menu_rolling + 2
+        sta.w menu_rolling + 4
+        sta.w menu_rolling + 6
+        sta.w menu_rolling + 8
+        sta.w menu_rolling + 10
+        plp
+        jsr.l reset_sprites_trampoline
+        rtl
+    ; swap_redraw_hook_impl_body
+    ; Called after item swap to redraw visible items correctly.
+    ; Must render to the correct circular buffer slots based on current buffer_pos.
+    ; Does NOT reset buffer_pos - we stay at the current scroll position.
+
+    swap_redraw_hook_impl_body:
+    """Field profile: post-swap re-render of all 11 slots via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1B1A
+        ldx.w #menu_rolling
+        jsr.l rolling_engine.rolling_engine_swap_redraw
+        plp
+        rtl
+
+
+    ; _clear_inventory_slot
+    ; Clears a single inventory slot in the tilemap buffer.
+    ; Input: menu_rolling_slot_index = slot to clear (0-10)
+    ; Used when item index is out of bounds (>= 48)
+
+    _clear_inventory_slot:
+        php
+        phb
+    ; Set data bank to $7E for WRAM access
+        lda #0x7E
+        pha
+        plb
+    ; Save registers - use 16-bit mode for consistent push/pop
+        rep #0x30
+    ; 16-bit A and X/Y
+        pha
+        phx
+        phy
+        lda.b 0x29
+        pha
+    ; Set $29 = $B600 for BG1 tilemap buffer
+        lda.w #0xB600
+        sta.b 0x29
+    ; Calculate Y = slot_index * 128 + 70
+        lda.w menu_rolling_slot_index
+        and.w #0x00FF
+        xba
+    ; A = slot * 256
+        lsr
+    ; A = slot * 128
+        clc
+        adc.w #0x0044
+    ; + 68 (64 border + 4 margin)
+        tay
+    ; Y = tilemap offset (16-bit)
+        sep #0x20
+    ; 8-bit A for tile writes (X/Y stay 16-bit)
+    ; Clear the item name area (12 tiles = 24 bytes)
+    ; Use $FF as blank tile
+    ; Note: X is 16-bit but we only use low byte; loop works correctly
+        ldx.w #12
+    ; 12 tiles for item name + quantity (force 16-bit immediate)
+
+    _clear_slot_loop:
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+        iny
+        lda #0x04
+    ; Palette 4 (matches normal items)
+        sta (0x29), y
+        iny
+        dex
+        bne _clear_slot_loop
+    ; Restore registers - must match push mode
+        rep #0x20
+    ; 16-bit A for pop (X/Y already 16-bit)
+        pla
+        sta.b 0x29
+        ply
+        plx
+        pla
+        plb
+        plp
+        rts
+    ; _clear_trash_area
+    ; Clears the 2x2 trash can area with blank tiles ($FF).
+    ; Called before drawing normal items to remove any leftover trash icon.
+    ; Input: Y = tilemap offset
+    ;        ($29) = tilemap base
+    ;
+
+    _clear_trash_area:
+        phy
+    ; Save Y
+    ; First row: 2 tiles
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+    ; Second row: Y + 64 from start
+        ply
+    ; Restore original Y
+        phy
+    ; Save again
+        rep #0x20
+        tya
+        clc
+        adc.w #64
+        tay
+        sep #0x20
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+        ply
+    ; Restore Y
+        rts
+
+    draw_trash_single_column:
+
+
+    """
+    Draws the trash can 2x2 tile graphic for single-column inventory.
+    Input: Y = tilemap offset (from slot calculation)
+    ($29) = tilemap base ($B600)
+    menu_rolling_slot_index = current slot
+    Tiles: $04 (top-left), $05 (top-right), $06 (bottom-left), $07 (bottom-right)
+
+    Tilemap format: [tile_number, attributes] pairs
+    Each row is 64 bytes (32 tiles × 2 bytes)
+    """
+
+
+    ; Y points to start of item slot area
+    ; Draw 2x2 trash can icon, then clear remaining 10 tiles per row
+    ; Save starting Y for second row calculation
+        rep #0x20
+    ; 16-bit for push
+        phy
+    ; Save starting Y
+        sep #0x20
+    ; 8-bit A
+    ; First row: tiles $04, $05
+        lda #0x04
+    ; Tile $04 (top-left of trash can)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute: palette 0, no flip
+        sta (0x29), y
+        iny
+        lda #0x05
+    ; Tile $05 (top-right)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+    ; Clear remaining 13 tiles on first row (10 name + colon + 2 digits)
+        ldx.w #13
+
+    _clear_row1:
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+        dex
+        bne _clear_row1
+    ; Restore starting Y and add 64 for second row
+        rep #0x20
+    ; 16-bit A
+        pla
+    ; Get starting Y
+        clc
+        adc.w #64
+    ; +64 bytes = next tilemap row
+        tay
+        sep #0x20
+    ; 8-bit A
+    ; Second row: tiles $06, $07
+        lda #0x06
+    ; Tile $06 (bottom-left)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+        lda #0x07
+    ; Tile $07 (bottom-right)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+    ; Clear remaining 13 tiles on second row (10 name + colon + 2 digits)
+        ldx.w #13
+
+    _clear_row2:
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+        dex
+        bne _clear_row2
+        rts
+    ; NmiDmaTransferCheck_Impl moved to bank $20 (battle/inventory_rolling.s)
+    ; as field_menu_nmi_dma_transfer_check_impl to save space in bank $01
+    ; check_and_clear_count_impl
+    ; Called from DrawItemSlot to check item ID and handle count display.
+    ; - If item ID is 0: writes $FF tiles to clear count, skips to RTS
+    ; - If item ID is $FE: skips to RTS (no clearing needed)
+    ; - Otherwise: returns normally to draw count
+    ;
+    ; Input: $5a = pointer to item data, Y = tilemap offset, $29 = tilemap ptr
+    ; Modifies: A, return address on stack if skipping
+
+    check_and_clear_count_impl:
+    """Drop the count column for empty/used slots."""
+        lda (0x5a)
+    ; Load item ID
+        beq _clear_count
+    ; If 0, clear and skip
+        cmp #0xFE
+    ; Check for special item $FE
+        bne _normal_return
+    ; If not $FE, return normally to draw count
+    ; Item is $FE - skip count but don't clear
+        bra _skip_to_rts
+
+    _clear_count:
+    ; Write $FF (blank tiles) to count area: colon + 2 digits = 3 tiles
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+    ; Colon position
+        iny
+        lda.b 0xdb
+    ; Attribute byte
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+    ; First digit position
+        iny
+        lda.b 0xdb
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+    ; Second digit position
+        iny
+        lda.b 0xdb
+        sta (0x29), y
+
+    _skip_to_rts:
+    ; Replace the entire DrawItemSlot return chain so we land on the RTS at $01:A222
+    ; instead of falling back to $A205 (NOPs + colon write).
+    ;
+    ; Stack on entry to Impl (pushed by trampoline + caller, top first):
+    ;   [JSL Impl return: PCL, PCH, PB ($01)]
+    ;   [JSR check_and_clear_count return: PCL, PCH ($A204)]
+    ;   [DrawItemSlot caller's return ...]
+    ;
+    ; Pop both the JSL Impl return (3 bytes) AND the JSR trampoline return (2 bytes),
+    ; then push a synthetic return to $01:A222 (RTS) for our RTL.
+        pla  ; PCL of JSL Impl return
+        pla  ; PCH of JSL Impl return
+        pla  ; PB  of JSL Impl return
+        pla  ; PCL of JSR check_and_clear_count return
+        pla  ; PCH of JSR check_and_clear_count return
+    ; Stack top now = DrawItemSlot's own caller return.
+    ; Push return for RTL: $01:A221 (so RTL lands PC=$01:A222 = RTS).
+        lda #0x01
+        pha
+        lda #0xA2
+        pha
+        lda #0x21
+        pha
+
+    _normal_return:
+        rtl
+
+    circular_slot_calc:
+
+
+    """
+    Calculate tilemap Y offset using circular buffer position.
+    Called from patched code at $A1BA via CircularSlotCalc_ext.
+
+    Input: $5D = game's slot counter (0, 2, 4, 6... incremented by 2 per row)
+    $5A = src pointer ($1440 = inventory, $FF28 = treasure drops)
+    Output: Y = tilemap offset for circular buffer slot
+    Preserves: 16-bit A mode on exit
+    """
+
+
+        sep #0x20
+    ; 8-bit A
+    ; Drops list at $FF28 reuses _a181 but doesn't have any rolling state.
+    ; The original 2-col Y calc collides slots after we forced col=0 globally
+    ; via the `and #$00` patch at DrawItemSlot, so two drops would render at
+    ; the same scanline. Detect drops via $5B high byte and emit a unique
+    ; per-slot Y (slot * 64 + 4) instead of any circular math.
+        lda.b 0x5b
+        cmp #0xFF
+        bne _circ_slot_not_drops
+        lda.b 0x5d
+        rep #0x20
+        and.w #0x00FF
+        asl
+        asl
+        asl
+        asl
+        asl
+        asl  ; * 64. $5D increments by 2 → step = 128 bytes = 2 tilemap rows,
+    ; matching the 16-px-per-item layout original uses for col-0
+    ; drops with BG1VOFS=-32 ($93) pushing them down to scanline 32.
+        clc
+        adc.w #0x0044  ; col 2 + 1 tilemap row down (+$40) so the first slot
+    ; clears the window header before BG1VOFS shifts it.
+        tay
+        rts
+    _circ_slot_not_drops:
+    ; Inventory in treasure context uses treasure_rolling_buffer_pos.
+        lda.l 0x7E0000 + 0x9C06  ; treasure_hdma_enable (treasure_rolling + 0x06)
+        beq _circ_check_field
+        lda.b 0x5d
+        lsr
+        clc
+        adc.l 0x7E0000 + 0x9C01  ; treasure_rolling_buffer_pos (treasure_rolling + 0x01)
+    _t_circ_mod:
+        cmp #6  ; TREASURE_BUFFER_SLOTS
+        bcc _t_circ_done
+        sec
+        sbc #6
+        bra _t_circ_mod
+    _t_circ_done:
+        rep #0x20
+        and.w #0x00FF
+        xba
+        lsr
+        clc
+        adc.w #0x0004
+        tay
+        rts
+    _circ_check_field:
+    ; Check if circular buffer mode is active (HDMA enabled)
+        lda.l 0x7E0000 + menu_hdma_enable
+        beq _circ_slot_original
+    ; Not active, use original calculation
+    ; Circular buffer Y calculation
+    ; NOTE: Game increments $5D by 2 for each row (0, 2, 4, 6, 8, 10, 12, 14, 16, 18)
+    ; We must divide by 2 first to get the visual slot (0-9)
+        lda.b 0x5d
+    ; Load slot counter (0, 2, 4...)
+        lsr
+    ; Divide by 2 to get visual slot (0-9)
+        clc
+        adc.l 0x7E0000 + menu_rolling_buffer_pos
+    ; Add buffer_pos
+
+    _circ_slot_mod:
+        cmp #MENU_BUFFER_SLOTS
+    ; >= 11?
+        bcc _circ_slot_done
+        sec
+        sbc #MENU_BUFFER_SLOTS
+    ; Subtract 11 to wrap
+        bra _circ_slot_mod
+
+    _circ_slot_done:
+    ; A = circular slot (0-10)
+        rep #0x20
+    ; 16-bit A
+        and.w #0x00FF
+    ; Clear high byte (force 16-bit immediate)
+        xba
+    ; A = slot * 256
+        lsr
+    ; A = slot * 128
+        clc
+        adc.w #0x0004
+    ; + 4 (margin only, $29 already includes border)
+        tay
+    ; Y = tilemap offset
+        rts
+
+    _circ_slot_original:
+    ; Original game calculation: Y = ($5D / 2) * 128 + 4
+        lda.b 0x5d
+        lsr
+    ; /2
+        rep #0x20
+    ; 16-bit A
+        and.w #0x00FF
+    ; Clear high byte (force 16-bit immediate)
+        xba
+    ; *256
+        lsr
+    ; *128
+        clc
+        adc.w #0x0004
+    ; +4
+        tay
+        rts
+
+    circular_slot_calc_ext:
+    """Trampoline to call CircularSlotCalc from bank $01 patch at $A1BA"""
+        jsr.w circular_slot_calc
+        rtl
+    }
 }
diff --git a/src/ingame/inventory_rolling_trampolines.s b/src/ingame/inventory_rolling_trampolines.s
index 8463c9b..ae43c03 100644
--- a/src/ingame/inventory_rolling_trampolines.s
+++ b/src/ingame/inventory_rolling_trampolines.s
@@ -4,6 +4,8 @@ wrappers around original bank-$01 helpers used by the rolling code.
 """
 .include "src/ingame/macros.i"
 
+.extern items_menu_vwf.draw_field_item_name
+
 ;; Bank-$01 trampolines for inventory rolling routines living in bank $21.
 ;; Reclaimed space: $01:EBD2 onwards (from init_bg_scroll_hdma relocation).
 ;; Each trampoline = jsr.l + rts = 5 bytes.
@@ -17,7 +19,7 @@ wrappers around original bank-$01 helpers used by the rolling code.
 }
 
 .if INVENTORY_ROLLING_BUFFER {
-.alloc bank01_inventory_trampolines in bank01_trampolines {
+    .alloc bank01_inventory_trampolines in bank01_trampolines {
 check_and_clear_count:
 """Bank-$01 trampoline: bridge to `check_and_clear_count_impl` in bank $21."""
     jsr.l check_and_clear_count_impl
@@ -136,8 +138,30 @@ reset_sprites_trampoline:
     jsr 0x8D6A
 ; original @ $01:8D6A
     rtl
-}  ; end .alloc bank01_inventory_trampolines
-}  ; end .if INVENTORY_ROLLING_BUFFER
+
+draw_field_item_name_trampoline:
+"""
+Bank-$01 JSR-callable wrapper around the bank-20
+`items_menu_vwf.draw_field_item_name` RTL helper. The vanilla
+`DrawItemName` patch at $01:9060 / `DrawEquipItemName` at $01:9013
+must stay 4 bytes (jsr.w + rts) so the very next byte at $01:9064
+keeps holding the vanilla sprite-render sub-routine's `phx` opcode :
+overrunning into $01:9064 with a JSL + RTS (5 bytes) replaced that
+phx with the wrapper's RTS, which short-circuited the save-selection
+sprite path's `jsr $9064` and parked every sprite off-screen on the
+title-screen-press-A entry. Use this trampoline so the call site
+stays at 3+1 bytes and the vanilla sprite routine is preserved.
+"""
+
+
+    jsr.l items_menu_vwf.draw_field_item_name
+    rts
+    }
+
+
+; end .alloc bank01_inventory_trampolines
+}
+; end .if INVENTORY_ROLLING_BUFFER
 
 ;; Bank-$01 thunks + wrappers for the treasure exchange rolling buffer.
 ;; Mirror the field-menu set above but call the treasure_-prefixed bodies
@@ -145,7 +169,7 @@ reset_sprites_trampoline:
 ;; screen so the HDMA channel + tilemap buffer + WRAM shadow tables are
 ;; reused; only the per-menu state RAM differs.
 .if TREASURE_INVENTORY_ROLLING {
-.alloc bank01_treasure_trampolines in bank01_trampolines {
+    .alloc bank01_treasure_trampolines in bank01_trampolines {
 _treasure_check_and_clear_count:
     jsr.l treasure_check_and_clear_count_impl
     rts
@@ -195,8 +219,13 @@ treasure_scroll_down_trigger:
 """Treasure profile: input-driven scroll-down trigger."""
     lda.w treasure_scroll_state
     bne _t_down_abort
+    lda.w treasure_scroll_cooldown
+    bne _t_down_abort
     jsr.w _treasure_force_hdma_setup
     jsr.w _treasure_start_scroll_down
+    sep #0x20
+    lda.b #TREASURE_SCROLL_COOLDOWN_FRAMES
+    sta.w treasure_scroll_cooldown
     rts
 _t_down_abort:
     dec.w 0x1BB7
@@ -206,8 +235,13 @@ treasure_scroll_up_trigger:
 """Treasure profile: input-driven scroll-up trigger."""
     lda.w treasure_scroll_state
     bne _t_up_abort
+    lda.w treasure_scroll_cooldown
+    bne _t_up_abort
     jsr.w _treasure_force_hdma_setup
     jsr.w _treasure_start_scroll_up
+    sep #0x20
+    lda.b #TREASURE_SCROLL_COOLDOWN_FRAMES
+    sta.w treasure_scroll_cooldown
     rts
 _t_up_abort:
     inc.w 0x1BB7
@@ -271,6 +305,17 @@ by calling the original $82C0 so original per-frame work still runs.
 """
 
 
+; Cooldown tick : runs every popup frame regardless of button state
+; so the held-DOWN debounce drains uniformly. Without this the
+; trigger path's per-call dec only fired while DOWN was held, which
+; let auto-repeat consume the cooldown in 2-3 frames and scrolled
+; the treasure inv two items per visible tap.
+    sep #0x20
+    lda.w treasure_scroll_cooldown
+    beq _t_main_cd_done
+    dec.w treasure_scroll_cooldown
+
+_t_main_cd_done:
     lda.w treasure_scroll_state
     beq _treasure_main_check_drops_tick
     jsr.w _treasure_update_scroll_frame
@@ -423,5 +468,63 @@ the bottom border at screen y=112..120 (just below the 5th item).
 
 
     menu_window(0, 0, 30, 12)
-}  ; end .alloc bank01_treasure_trampolines
-}  ; end .if TREASURE_INVENTORY_ROLLING
+    }
+
+
+; end .alloc bank01_treasure_trampolines
+}
+; end .if TREASURE_INVENTORY_ROLLING
+
+.alloc bank01_shop_trampolines in bank01_trampolines {
+shop_quantity_text_hook:
+"""
+Render the owner welcome ('Que désirez vous ?') through the small-VWF
+description region, then fall through to the vanilla menu-text engine
+for the rest of the quantity block (Quantité + initial '1' digit).
+
+Called in place of the original `jsr $8301` at the buy ($01:C442) and
+sell ($01:C7E7) entry points  ; the matching `LDY #shops.quantity` at
+$01:C43F / $01:C7E4 is left in place so an unrelated future caller
+could still chain into $8301 with the slimmed block.
+"""
+
+
+    ldy.w #shops.que_desirez_vous
+    jsr.l items_description.draw_trampoline_pos
+    ldy.w #shops.quantity - 0x8000
+    jsr.w 0x8301  ; draw text at position (= display_text_in_menus thunk)
+    rts
+
+shop_welcome_text_hook:
+"""
+Render the shop owner's greeting ('Puis-je vous aider ?') through the
+small-VWF description region, then tail-jump to the vanilla menu-text
+engine for the remaining `Achat Vente Sortir` action labels.
+
+Called in place of the original `jmp $8301` at $01:C353. The matching
+`LDY #shops.welcome_and_actions` at $01:C350 is left in place  ; the
+slimmed `welcome_and_actions` block now holds only the action line.
+"""
+
+
+    ldy.w #shops.puis_je_vous_aider
+    jsr.l items_description.draw_trampoline_pos
+    ldy.w #shops.welcome_and_actions - 0x8000
+    jmp.w 0x8301  ; tail-call to draw positioned text
+
+shop_thanks_text_hook:
+"""
+Draw the thank-you window via the vanilla `$82FB` (which expects an
+empty trailing text block  ; see `thank_you_window` data), then render
+the actual 'Merci !' copy through the small-VWF description region.
+
+Called in place of the original `jsr $82FB` at $01:C751. The matching
+`LDY #shops.thank_you_window` at $01:C74E is left in place.
+"""
+
+
+    jsr.w 0x82FB  ; draw window + (empty) text
+    ldy.w #shops.merci
+    jsr.l items_description.draw_trampoline_pos
+    rts
+}
diff --git a/src/ingame/items_menu.s b/src/ingame/items_menu.s
index f9e60f4..413e849 100644
--- a/src/ingame/items_menu.s
+++ b/src/ingame/items_menu.s
@@ -1,16 +1,23 @@
 """
-Item-name expansion for the field-menu (mirror of `battle/items_patches.s`): mul-by-9 → mul-by-12 stride
-changes, $0F8000 → `assets_items_dat` pointer remaps.
+Item-name expansion for the field-menu (mirror of `battle/items_patches.s`): mul-by-9 -> mul-by-17 stride
+changes, $0F8000 -> `assets_items_unleashed_dat` pointer remaps.
+Field / drops / treasure rolling inventory all defer to vanilla
+DrawItemSlot at $01:9000, so patching here switches them in one
+go.
 """
+
+.extern items_menu_vwf.draw_field_item_name
 ; Item name expansion for menu system
-; Patches the multiply-by-9 to multiply-by-12
-; Also redirects $0F8000 references to assets_items_dat
+; Patches the multiply-by-9 to multiply-by-17
+; Also redirects $0F8000 references to assets_items_unleashed_dat
 
 ; --- Patch loop counter ---
 ; Original: 01/903F: A9 08  LDA #$08
+; Name field is ITEM_UNLEASHED_TEXT_SIZE chars in items_unleashed
+; (records are ITEM_UNLEASHED_RECORD_SIZE bytes = symbol + name).
 
 *=0x01903F
-    lda #0x0b
+    lda #ITEM_UNLEASHED_TEXT_SIZE
 
 ; --- Patch multiply logic ---
 
@@ -19,7 +26,7 @@ changes, $0F8000 → `assets_items_dat` pointer remaps.
 ; New: JSL to relocated routine (4 bytes) + 3 NOPs
 
 *=0x019023
-    jsr.l multiply_item_index_12
+    jsr.l multiply_item_index_17
     nop
     nop
     nop
@@ -31,26 +38,51 @@ changes, $0F8000 → `assets_items_dat` pointer remaps.
 ; Original: 01/902E: BF 00 80 0F  LDA $0F8000,X
 
 *=0x01902E
-    lda.l assets_items_dat, x
+    lda.l assets_items_unleashed_dat, x
 
 ; --- menu: item name (in loop) ---
 ; Original: 01/9043: BF 00 80 0F  LDA $0F8000,X
 
 *=0x019043
-    lda.l assets_items_dat, x
+    lda.l assets_items_unleashed_dat, x
+
+; ===== DRAWITEMNAME JSL HOOKS =====
+; Both vanilla entry points relocate to `items_menu_vwf.draw_field_item_name`
+; in bank $20. Initial stub mirrors the vanilla fixed-font body so the
+; visible output stays identical; subsequent phases will swap in the
+; small_vwf glyph blit + per-slot CHR allocator.
+
+; DrawEquipItemName ($01:9013): vanilla `phy ; phx ; lda ($60),y ; bra _9017`.
+; Route through the bank-01 jsr.w trampoline so the patch stays inside
+; the vanilla 4-byte slot and the byte at $01:9017 stays untouched.
+
+*=0x019013
+    lda (0x60), y
+    jsr.w draw_field_item_name_trampoline
+    rts
+
+; DrawItemName ($01:9060): vanilla `phy ; phy ; bra _9017` -> caller already
+; passed A = item_id. Use the bank-01 jsr.w trampoline (3 bytes) + rts
+; (1 byte) so the FIVE-byte JSL.L + RTS no longer spills into the
+; sprite-render sub-routine at $01:9064 (the save-screen sprite path
+; jsr's $9064 directly).
+
+*=0x019060
+    jsr.w draw_field_item_name_trampoline
+    rts
 
 ; ===== COLON/QUANTITY POSITION PATCHES =====
-; Item names expanded from 9 to 12 bytes (+3 chars = +6 VRAM bytes)
-; Change offset from $0052 to $0058
+; Item names expanded from 9 to 16 bytes (+7 chars = +14 VRAM bytes)
+; Change offset from $0052 to $0060
 
 ; --- DrawItemSlot: left column colon/qty position ---
 ; Original: 01/A1FC: 69 52 00  ADC #$0052
 
 *=0x01A1FC
-    adc #0x0058
+    adc #0x0060
 
 ; --- DrawItemSlot: right column colon/qty position ---
 ; Original: 01/A236: 69 52 00  ADC #$0052
 
 *=0x01A236
-    adc #0x0058
+    adc #0x0060
diff --git a/src/ingame/items_menu_vwf.s b/src/ingame/items_menu_vwf.s
new file mode 100644
index 0000000..542af0c
--- /dev/null
+++ b/src/ingame/items_menu_vwf.s
@@ -0,0 +1,345 @@
+"""
+Field-menu item-name VWF renderer.
+
+Replaces vanilla `DrawItemName` / `DrawEquipItemName` at $01:9013 /
+$01:9060. The vanilla routine wrote 16-pixel-tall fixed-font tile
+ids across two tilemap rows ($29) (dakuten / top half) and
+($1D = $29 + $40) (kana / bottom half). We keep the 16-pixel-tall
+slot layout but render the name as 8-pixel-tall variable-width
+glyphs in the BOTTOM tile row only  ; the TOP row is filled with
+blank ($FF) tiles + the menu palette byte ($DB) so the slot still
+takes two tilemap rows of vertical space.
+
+Inputs (caller convention preserved):
+  A   = item id (8-bit), set by either entry point.
+  Y   = tilemap byte offset within the row (set by the rolling
+        engine before each row render).
+  $29 = top-row tilemap pointer (16-bit, in WRAM bank $7E).
+  $5D = slot index 0..N-1 (set by `_menu_render_item_to_slot` in
+        `src/ingame/inventory_rolling.s` before each per-row call).
+  $DB = palette / attribute byte for fixed cells.
+
+Outputs:
+  Top row at ($29),y..($29),y+31     filled with $FF + ($DB|$34).
+  Bottom row at ($29+$40),y..        filled with VWF glyphs +
+                                     palette via small_vwf.
+  X   = `items_unleashed` offset for the last byte consumed.
+
+Tile-id allocation:
+  Each visible slot owns ITEM_VWF_TILE_BUDGET (=10) tile ids
+  starting at ITEM_VWF_TILE_BASE + slot * budget, same scheme the
+  battle inventory uses. The render_allocator clamp at
+  slot_limit_low keeps overflow contained to the slot.
+
+Status:
+  Wiring scaffold. The bank-20 entry point is in place and the
+  $01:9013 / $01:9060 hooks JSL into it, but the body still falls
+  back to the fixed-font path so the visible output matches what
+  vanilla DrawItemName produced before the switch. Subsequent
+  phases will swap the loop body to `small_vwf.render.display_char`
+  + per-slot CHR clear + top-row blanking.
+"""
+
+
+.include "src/items.i"
+.include "../bank20.i"
+.include "src/battle/inventory_budget.i"
+
+.include "src/libmz.i"
+.include "src/vwf_state.i"
+
+.extern assets_items_unleashed_dat
+.extern render
+.extern render.render_with_config
+.extern wait_for_vblank_long
+.extern dma_transfer_to_vram
+
+
+.alloc items_menu_vwf_block in bank20_reloc {
+    .scope items_menu_vwf {
+        """
+    VWF render path for field-menu / treasure / drops item-name slots.
+    Hijacks vanilla `DrawItemName` ($01:9060) to populate VwfConfig at
+    VWF_CONFIG_BASE and dispatch through `render.render_with_config`, so
+    the patched menu chrome keeps the slot layout while the glyphs go
+    through the variable-width font engine.
+        """
+    draw_field_item_name:
+    """
+    Bank-20 field-menu item-name render driven by VwfConfig.
+
+    Stages the item name in `VWF_TEXT_BUFFER` with a $00 terminator,
+    fills `VWF_CONFIG_BASE` with per-slot tile budget + tilemap dest,
+    fills the top tilemap row with $FF blanks (so the 16-pixel-tall
+    slot keeps its height), writes the items_unleashed symbol byte
+    + palette to the bottom row's first tile, then calls the unified
+    `render.render_with_config` to blit the rest. Field uses K=5
+    tiles/slot (FIELD_ITEM_VWF_TILE_BUDGET) so 11 slots fit inside the
+    $C0..$F6 tile-id window.
+    """
+
+
+        sta.b 0x43  ; item id
+        php
+        sep #0x20
+        rep #0x10
+    ; Preserve caller's Y. Vanilla `DrawEquipItemName` (`$01:9013`) and
+    ; `DrawItemName` (`$01:9060`) each phy at entry and ply before rts ;
+    ; the caller `DrawItemSlot` continues with `tya ; adc #$60 ; tay`
+    ; to position the colon glyph, so Y must come back unchanged.
+        phy
+    ; --- X = item id * ITEM_UNLEASHED_RECORD_SIZE (inline mul-by-17) ---
+        rep #0x20
+        lda.b 0x43
+        and.w #0x00FF
+        pha
+        asl
+        asl
+        asl
+        asl  ; * 16
+        clc
+        adc 0x01, s  ; + id = * 17
+        tax
+        pla
+        sep #0x20
+    ; --- Copy items_unleashed name bytes into VWF_TEXT_BUFFER, terminate $00 ---
+    ; Byte 0 of the record is the symbol (rendered separately as fixed
+    ; tile below) ; bytes 1..ITEM_UNLEASHED_TEXT_SIZE go into the buffer.
+    ; X is the destination index (only abs,x works with sta.l) ; the
+    ; source offset rides in long SRAM scratch `VWF_SRC_OFFSET` so we
+    ; do not steal a direct-page byte from vanilla's menu loop (the
+    ; original placement on DP $45 clashed with the items code's row
+    ; counter and broke the per-slot copy).
+        rep #0x20
+        txa
+        inc  ; skip symbol byte
+        sta.l VWF_SRC_OFFSET
+        sep #0x20
+        ldx.w #0x0000
+
+    _copy_loop:
+        phx
+        rep #0x20
+        lda.l VWF_SRC_OFFSET
+        tax
+        sep #0x20
+        lda.l assets_items_unleashed_dat, x
+        pha  ; save the byte so the 16-bit src-pointer update does not clobber it
+        inx
+        rep #0x20
+        txa
+        sta.l VWF_SRC_OFFSET
+        sep #0x20
+        pla
+        plx
+        sta.l VWF_TEXT_BUFFER, x
+        inx
+        cpx.w #ITEM_UNLEASHED_TEXT_SIZE
+        bne _copy_loop
+        lda.b #0x00
+        sta.l VWF_TEXT_BUFFER, x  ; null terminator
+    ; --- Populate VwfConfig.tile_id_base = FIELD base + $5D * K ---
+    ; K = FIELD_ITEM_VWF_TILE_BUDGET (=10). slot * 10 = slot*8 + slot*2.
+    ; Store the full 16-bit value: slots 6..10 produce tile_id_base
+    ; $C0 + 10*K = $C0 + 100 = $124 which the 8-bit allocator wrapped
+    ; back into the menu font CHR range. The wide init keeps tile_id
+    ; bit 8 (which `tilemap_write_no_inc` ORs in via $01 on the high
+    ; tilemap byte) intact.
+        rep #0x20
+        lda.b 0x5D
+        and.w #0x00FF
+        pha
+        asl
+        asl
+        asl  ; * 8
+        clc
+        adc 0x01, s  ; + slot = * 9
+        clc
+        adc 0x01, s  ; + slot = * 10
+        clc
+        adc.w #FIELD_ITEM_VWF_TILE_BASE
+        sta.l VWF_CONFIG_BASE + VwfConfig.tile_id_base  ; word
+        pla  ; balance
+        sep #0x20
+        lda.b #FIELD_ITEM_VWF_TILE_BUDGET
+        sta.l VWF_CONFIG_BASE + VwfConfig.slot_budget
+    ; CHR -> VRAM flush descriptor. Menu PPU runs in Mode 0
+    ; (BGMODE = $00 at `ff4decomp/menu/menu.asm:3878`) with BG34NBA = $22
+    ; -> BG3 CHR at VRAM word $2000 (byte $4000). Two descriptors live in
+    ; SRAM so the NMI flush can DMA treasure (primary) and drops (secondary)
+    ; in disjoint slices without one panel's range trampling the other's
+    ; through a single combined DMA. VWF_CALLER_CTX (set by drops_rolling
+    ; around its vanilla JSR chain) picks which descriptor this call's
+    ; flush targets ; tile_id_base / slot_budget / tilemap_base / flags
+    ; stay primary regardless since they are per-call render inputs, not
+    ; per-region flush params.
+    ; Tight `== 1` check so stale SRAM from old savestates (or any future
+    ; CTX value not yet defined) falls back to the primary path instead
+    ; of misrouting to drops. Without this, savestates captured before
+    ; the secondary descriptor existed carried random bytes at
+    ; VWF_CALLER_CTX and any non-zero value sent the very first render
+    ; into the drops branch on the wrong inventory.
+        sep #0x20
+        lda.l VWF_CALLER_CTX
+        cmp.b #0x01
+        beq _write_secondary_desc
+        rep #0x20
+    ; --- Primary descriptor (treasure / field-items / default) ---
+        lda.w #FIELD_VWF_VRAM_DEST_WORD
+        sta.l VWF_CONFIG_BASE + VwfConfig.chr_vram_word
+        lda.w #FIELD_VWF_PRIMARY_BYTE_COUNT
+        sta.l VWF_CONFIG_BASE + VwfConfig.chr_byte_count
+        bra _desc_done
+
+    _write_secondary_desc:
+    ; --- Secondary descriptor (drops in treasure popup) ---
+        rep #0x20
+        lda.w #DROPS_VWF_VRAM_DEST_WORD
+        sta.l VWF_CHR_VRAM_WORD_B
+        lda.w #DROPS_VWF_BYTE_COUNT
+        sta.l VWF_CHR_BYTE_COUNT_B
+        lda.w #DROPS_VWF_CHR_SRC_OFFSET
+        sta.l VWF_CHR_SRC_OFFSET_B
+
+    _desc_done:
+        sep #0x20
+    ; Tilemap attr OR mask. Field VWF tile_ids live in the 9-bit
+    ; window ($100..$169) so bit 0 of the attr byte (= tile_id bit 8)
+    ; must be set ; the allocator stores $0100+, the tilemap entry
+    ; writes the low byte to the tile_id slot and ORs `$01` into the
+    ; attr byte to give the PPU the full 9-bit tile_id.
+        lda.b #0x01
+        sta.l VWF_CONFIG_BASE + VwfConfig.flags
+    ; --- VwfConfig.tilemap_base = $29 + $40 + Y + 2 (skip symbol slot) ---
+    ; Caller's Y is the 16-bit byte offset of the top row tile we are
+    ; about to write the symbol into ; VWF chars start two bytes later.
+    ; X-flag is 16-bit (we did rep #$10 at entry) so `tya` returns the
+    ; full Y. An earlier version of this helper masked Y to its low
+    ; byte with `and #$00FF`, which made every slot past slot 1
+    ; collapse onto slot 0's tilemap base (slot 2's Y = $0144 -> $44).
+        rep #0x20
+        tya
+        clc
+        adc.b 0x29
+        clc
+        adc.w #0x0042  ; + $40 (next row) + $02 (past symbol)
+        sta.l VWF_CONFIG_BASE + VwfConfig.tilemap_base
+        sep #0x20
+    ; --- Top row: $FF tile + palette across the full slot width
+    ; (1 symbol + ITEM_UNLEASHED_TEXT_SIZE name + trailing blanks fit
+    ; into the same Y window the vanilla loop walked) ---
+        ldx.w #0x0000
+
+    _top_loop:
+        lda.b #0xFF
+        sta (0x29), y
+        iny
+        lda.b 0xDB
+        ora.b 0x34
+        sta (0x29), y
+        iny
+        inx
+        cpx.w #( 1 + ITEM_UNLEASHED_TEXT_SIZE )
+        bne _top_loop
+    ; --- Bottom row: $FF blank pre-fill across the slot width so any
+    ; tilemap entries past the new glyph count stop showing PREVIOUS
+    ; frame's tile_ids. display_char rewrites the front-of-row entries
+    ; as it allocates new tile_ids per glyph ; the leftover positions
+    ; stay $FF blank instead of pointing at stale CHR (e.g. before this
+    ; pre-fill, drops slot 1's bottom row kept references to slot 0's
+    ; tile range whenever the previous frame had a longer name, which
+    ; rendered the start of "Aiguille d'or" inside the next slot's row
+    ; as soon as drops + treasure stopped sharing a single tile region).
+    ;
+    ; Y enters this block at Y_orig + (1+ITEM_UNLEASHED_TEXT_SIZE)*2,
+    ; pointing past the top-row run ; advance to the bottom-row start
+    ; (next BG row = +$40 from top-row base = +$40 - run_size from
+    ; current Y), blank 17 cells, then restore Y back to Y_orig for the
+    ; symbol-write block below.
+        rep #0x20
+        tya
+        clc
+        adc.w #( 0x0040 - ( 1 + ITEM_UNLEASHED_TEXT_SIZE ) * 2 )
+        tay
+        sep #0x20
+        ldx.w #0x0000
+
+    _bottom_blank_loop:
+        lda.b #0xFF
+        sta (0x29), y
+        iny
+        lda.b 0xDB
+        ora.b 0x34
+        sta (0x29), y
+        iny
+        inx
+        cpx.w #( 1 + ITEM_UNLEASHED_TEXT_SIZE )
+        bne _bottom_blank_loop
+    ; --- Restore caller's Y to point at the symbol slot, write symbol +
+    ; palette to the bottom row first tile. Y is currently
+    ; Y_orig + $40 + (1+ITEM_UNLEASHED_TEXT_SIZE)*2 after the bottom
+    ; blank ; subtract both terms to land back at Y_orig.
+        rep #0x20
+        tya
+        sec
+        sbc.w #( 0x0040 + ( 1 + ITEM_UNLEASHED_TEXT_SIZE ) * 2 )
+        tay
+        lda.b 0x29
+        clc
+        adc.w #0x0040
+        sta.b 0x1D
+        sep #0x20
+    ; X currently 0 from the top-row loop ; re-fetch items_unleashed offset.
+        rep #0x20
+        lda.b 0x43
+        and.w #0x00FF
+        pha
+        asl
+        asl
+        asl
+        asl
+        clc
+        adc 0x01, s
+        tax
+        pla
+        sep #0x20
+        lda.l assets_items_unleashed_dat, x
+        sta (0x1D), y  ; bottom-row symbol tile
+        iny
+        lda.b 0xDB
+        ora.b 0x34
+        sta (0x1D), y  ; bottom-row symbol palette
+        iny
+    ; --- Run the unified renderer over VWF_TEXT_BUFFER ---
+        jsr.l render_with_config_trampoline
+    ; render_with_config sets VWF_CHR_DIRTY=1 unconditionally. For drops
+    ; (CTX=1) ADDITIONALLY raise DIRTY_B so the NMI's secondary flush
+    ; covers drops's region this frame. Treasure's primary DIRTY must
+    ; stay set untouched : treasure rendered its own slots earlier in the
+    ; same frame and clearing here would skip its flush, leaking last-
+    ; frame's CHR back into the treasure inventory band. Leaving DIRTY=1
+    ; is harmless on drops-only frames: the primary DMA covers treasure
+    ; VRAM range only ($5000..$5400) which drops never writes into.
+        sep #0x20
+        lda.l VWF_CALLER_CTX
+        cmp.b #0x01
+        bne _dirty_done
+        lda.b #0x01
+        sta.l VWF_CHR_DIRTY_B
+
+    _dirty_done:
+        ply
+        plp
+        rtl
+
+    render_with_config_trampoline:
+    """
+    Trampoline that pops the bank then jsr's the bank-local entry
+    in small_vwf, paired RTL so callers stay long-jmp clean.
+    """
+
+
+        jsr.w render.render_with_config
+        rtl
+    }
+}
diff --git a/src/ingame/key_item_picker.s b/src/ingame/key_item_picker.s
index 9a89fc6..94e2d05 100644
--- a/src/ingame/key_item_picker.s
+++ b/src/ingame/key_item_picker.s
@@ -52,7 +52,11 @@ KEY_ITEM_SCROLL_LIMIT := 42
 KEY_ITEM_SCROLL_PIXELS_PER_FRAME := 8
 KEY_ITEM_SCROLL_TOTAL_PIXELS := 16
 
-key_item_rolling := 0x1BF0
+; Key-item picker state moved out of $1B00-$1BFF to clean $7E:9C60 for
+; the same reason as treasure ($9C00) + drops ($9C30) : engine path
+; needs 35 bytes per instance, $1B00-$1BFF is too small and vanilla
+; sprite code stomps past $1BEB.
+key_item_rolling := 0x9C60
 key_item_rolling_top_row := key_item_rolling + RollingBufferState.top_row
 key_item_rolling_buffer_pos := key_item_rolling + RollingBufferState.buffer_pos
 key_item_rolling_edge_row := key_item_rolling + RollingBufferState.edge_row
@@ -66,7 +70,7 @@ key_item_transfer_pending := key_item_rolling + RollingBufferState.transfer_pend
 key_item_scroll_anim_offset := key_item_rolling + RollingBufferState.scroll_anim_offset
 key_item_hdma_copy_pending := key_item_rolling + RollingBufferState.hdma_copy_pending
 
-key_item_scroll_pos := 0x1BFF
+key_item_scroll_pos := 0x9C8F
 
 KEY_ITEM_HDMA_TABLE_ADDR := 0x9900
 KEY_ITEM_HDMA_TABLE := 0x7E9900
@@ -84,323 +88,503 @@ KEY_ITEM_HDMA4_SRC_LO := 0x4342
 KEY_ITEM_HDMA4_SRC_BANK := 0x4344
 
 
-key_item_ensure_hdma_initialized:
-"""
-Lazy-capture $9F (BG3VOFS shadow) on first call, stash in base_scroll. HDMA channel enable deferred until the
-picker has its own window draw + visible loop wired — leaving ch4 enabled here corrupts the field BG3 layer
-(Cecil walks on a split screen) since the engine's RTL goes back to the field via EventCmd_f7 without any
-teardown.
-"""
-    rep #0x20
-    lda.w key_item_rolling_base_scroll
-    cmp.w #0xFFFF
-    bne _key_item_hdma_already_init
-    lda.l 0x7E019F
-    sta.w key_item_rolling_base_scroll
-    sep #0x20
-    jsr.w _key_item_init_hdma_channel
-    lda #KEY_ITEM_HDMA_CHANNEL_BIT
-    sta.l 0x7E1BAE
-    sta.w key_item_hdma_enable
-    rts
-
-_key_item_hdma_already_init:
-    sep #0x20
-    rts
-
-
-_key_item_init_hdma_channel:
-"""Configure HDMA ch4: DIRECT mode, dest BG3VOFS ($2112), source = picker shadow table at $7E:9900."""
-    php
-    sep #0x20
-    jsr.w update_key_item_scroll_hdma
-    lda #0x02
-    sta.l KEY_ITEM_HDMA4_CTRL
-    lda #0x12
-    sta.l KEY_ITEM_HDMA4_DEST
-    rep #0x20
-    lda.w #KEY_ITEM_HDMA_TABLE_ADDR
-    sta.l KEY_ITEM_HDMA4_SRC_LO
-    sep #0x20
-    lda #KEY_ITEM_HDMA_BANK
-    sta.l KEY_ITEM_HDMA4_SRC_BANK
-    plp
-    rts
-
-
-key_item_render_item_to_slot:
-"""Render filtered item from $7E:0712 + edge_row*Item.__size into BG3 buffer at $7E:D600 + slot_index*128 + 0x44."""
-    php
-    phb
-    lda #0x7E
-    pha
-    plb
-    rep #0x30
-    pha
-    phx
-    phy
-    lda.b 0x5a
-    pha
-    lda.b 0x29
-    pha
-    lda.b 0x45
-    pha
-    lda.b 0x33
-    pha
-    sep #0x20
-    lda.b 0x5d
-    pha
-    lda.b 0xDB
-    pha
-    rep #0x20
-    lda.w #0xD600
-    sta.b 0x29
-    sep #0x20
-    lda.w key_item_rolling_edge_row
-    asl
-    clc
-    adc #0x12
-    sta.b 0x5a
-    lda #0x07
-    adc #0x00
-    sta.b 0x5b
-    rep #0x20
-    lda.b 0x5a
-    tax
-    sep #0x20
-    lda.l 0x7E0000 + Item.id, x
-    pha
-    lda.l 0x7E0000 + Item.qty, x
-    sta.b 0x5C
-    stz.b 0x34
-    pla
-    jsr.l check_can_use_item_trampoline
-    lda.w key_item_rolling_slot_index
-    sta.b 0x5d
-    rep #0x20
-    lda.w key_item_rolling_slot_index
-    and.w #0x00FF
-    xba
-    lsr
-    clc
-    adc.w #0x0444
-    tay
-    sep #0x20
-    jsr.l draw_item_slot_inner_trampoline
-    pla
-    sta.b 0xDB
-    pla
-    sta.b 0x5d
-    rep #0x20
-    pla
-    sta.b 0x33
-    pla
-    sta.b 0x45
-    pla
-    sta.b 0x29
-    pla
-    sta.b 0x5a
-    rep #0x10
-    ply
-    plx
-    pla
-    plb
-    plp
-    rts
-
-
-key_item_render_all:
-"""
-Replacement for original UpdateItemText. Original just clears $0774 (text-buffer scratch) and walks $0712 to lay
-out 4x4 grid into the BG3 buffer at $7E:D600. We replace with engine_init_rolling_buffer which renders 6
-single-col slots from $0712 into the same buffer. Original NMI's BG3 transfer then pushes them to VRAM as part
-of the existing item-window flow ($EB=$01 latched by original preamble at $00:AF53).
-"""
-    php
-    rep #0x10
-    sep #0x20
-    jsr.l key_item_init_impl
-    ; engine's ensure_hdma turned $1BAE bit 4 on; clear so original NMI
-    ; doesn't try to drive HDMA we haven't fully wired (per-scanline
-    ; bands not yet matched to the picker rows).
-    lda #0x00
-    sta.l 0x7E1BAE
-    lda #0x00
-    sta.l 0x00420C
-    plp
-    rtl
-
-
-clear_key_item_slot:
-"""Blank one tilemap row at slot_index in the BG3 buffer."""
-    php
-    phb
-    lda #0x7E
-    pha
-    plb
-    rep #0x30
-    pha
-    phx
-    phy
-    lda.b 0x29
-    pha
-    lda.w #0xD600
-    sta.b 0x29
-    lda.w key_item_rolling_slot_index
-    and.w #0x00FF
-    xba
-    lsr
-    clc
-    adc.w #0x0044
-    tay
-    sep #0x20
-    ldx.w #0x0000
-
-_clear_key_loop:
-    lda #0x00
-    sta (0x29), y
-    iny
-    inx
-    cpx.w #0x0040
-    bne _clear_key_loop
-    rep #0x20
-    pla
-    sta.b 0x29
-    rep #0x10
-    ply
-    plx
-    pla
-    plb
-    plp
-    rts
-
-key_item_init_filter:
-"""
-Filter $1440 -> $0712. Faithful inline port of original InitItemList ($00:B2D5 in actual ROM, off-by-2 from
-ff4decomp notes). Clears the 96-byte filter buffer, walks 48 inventory items, copies (id, qty) pairs whose IDs
-are key items: [$CE..$E6] u [$EB..$FD].
-"""
-    php
-    phb
-    sep #0x20
-    rep #0x10
-    lda #0x7E
-    pha
-    plb
-    ldx.w #0x0000
-
-_filter_clear:
-    stz.w 0x0712, x
-    inx
-    cpx.w #0x0060
-    bne _filter_clear
-    ldx.w #0x0000
-    ldy.w #0x0000
-
-_filter_walk:
-    lda.w 0x1440, x
-    cmp #0xCE
-    bcc _filter_next
-    cmp #0xE7
-    bcc _filter_accept
-    cmp #0xEB
-    bcc _filter_next
-    cmp #0xFE
-    bcs _filter_next
-
-_filter_accept:
-    sta.w 0x0712, y
-    lda.w 0x1441, x
-    sta.w 0x0713, y
-    iny
-    iny
-
-_filter_next:
-    inx
-    inx
-    cpx.w #0x0060
-    bne _filter_walk
-    plb
-    plp
-    rts
-
-update_key_item_scroll_hdma:
-"""Build the key-item HDMA shadow table via the shared engine."""
-    engine_update_scroll_hdma(key_item_rolling, KEY_ITEM_HDMA_SHADOW, KEY_ITEM_BUFFER_SLOTS, KEY_ITEM_VISIBLE_ITEMS, _key_item_hdma_header, _key_item_hdma_footer, _key_item_hdma_signal)  ; noqa: E501
-
-_key_item_hdma_header:
-"""Picker HDMA header — 112 lines at BASE (top half = field map preserved)."""
-    sep #0x20
-    lda #112
-    sta.l KEY_ITEM_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w key_item_rolling_base_scroll
-    sta.l KEY_ITEM_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_key_item_hdma_footer:
-"""Picker HDMA footer — 16 lines at BASE+16 to hide prefetch slot."""
-    sep #0x20
-    lda #16
-    sta.l KEY_ITEM_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w key_item_rolling_base_scroll
-    clc
-    adc.w #16
-    sta.l KEY_ITEM_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_key_item_hdma_signal:
-"""NMI shadow-copy signal — set both picker copy_pending + shared $1BB6 mirror."""
-    sep #0x20
-    lda #0x01
-    sta.w key_item_hdma_copy_pending
-    rts
-
-
-key_item_init_impl:
-"""Init key-item picker (filter $1440 -> $0712, then engine init)."""
-    jsr.w key_item_init_filter
-    engine_init_rolling_buffer(key_item_rolling, KEY_ITEM_BUFFER_SLOTS, _key_item_draw_window, key_item_ensure_hdma_initialized, key_item_render_item_to_slot)  ; noqa: E501
-
-
-_key_item_draw_window:
-"""
-Picker is invoked from inside original ShowItemWindow which already drew the
-picker frame via its IRQ slide. No-op.
-"""
-    rts
-
-key_item_scroll_down_prepare:
-"""Picker scroll-down pre-render."""
-    engine_scroll_down_prepare(key_item_rolling, key_item_scroll_pos, KEY_ITEM_SCROLL_LIMIT, KEY_ITEM_VISIBLE_ITEMS, KEY_ITEM_BUFFER_SLOTS, key_item_ensure_hdma_initialized, key_item_render_item_to_slot, update_key_item_scroll_hdma)  ; noqa: E501
-
-key_item_scroll_up_prepare:
-"""Picker scroll-up pre-render."""
-    engine_scroll_up_prepare(key_item_rolling, key_item_scroll_pos, KEY_ITEM_BUFFER_SLOTS, key_item_ensure_hdma_initialized, key_item_render_item_to_slot, update_key_item_scroll_hdma)  ; noqa: E501
-
-key_item_start_scroll_down_impl:
-"""Picker: kick scroll-down state machine."""
-    engine_start_scroll_down(key_item_rolling, key_item_scroll_pos, KEY_ITEM_VISIBLE_ITEMS, KEY_ITEM_BUFFER_SLOTS, KEY_ITEM_SCROLL_TOTAL_PIXELS, KEY_ITEM_SCROLL_PIXELS_PER_FRAME, key_item_ensure_hdma_initialized, key_item_render_item_to_slot, update_key_item_scroll_hdma)  ; noqa: E501
-
-key_item_start_scroll_up_impl:
-"""Picker: kick scroll-up state machine."""
-    engine_start_scroll_up(key_item_rolling, key_item_scroll_pos, KEY_ITEM_BUFFER_SLOTS, KEY_ITEM_SCROLL_TOTAL_PIXELS, key_item_ensure_hdma_initialized, key_item_render_item_to_slot, update_key_item_scroll_hdma)  ; noqa: E501
-
-key_item_update_scroll_frame_impl:
-"""Picker: per-frame scroll animation tick."""
-    engine_update_scroll_frame(key_item_rolling, KEY_ITEM_SCROLL_PIXELS_PER_FRAME, update_key_item_scroll_hdma)
-
-key_item_finish_scroll_impl:
-"""Picker: end-of-animation pre-render + cleanup."""
-    engine_finish_scroll(key_item_rolling, key_item_scroll_pos, KEY_ITEM_VISIBLE_ITEMS, KEY_ITEM_BUFFER_SLOTS, KEY_ITEM_TOTAL_ITEMS, key_item_render_item_to_slot, update_key_item_scroll_hdma)  ; noqa: E501
-
-key_item_refresh_slots_impl:
-"""Picker: re-render all slots."""
-    engine_refresh_slots(key_item_rolling, key_item_scroll_pos, KEY_ITEM_BUFFER_SLOTS, key_item_render_item_to_slot)
+.include "../bank20.i"
+
+.alloc key_item_picker_block in bank20_reloc {
+    key_item_ensure_hdma_initialized:
+    """
+    Lazy-capture $9F (BG3VOFS shadow) on first call, stash in base_scroll. HDMA channel enable deferred until the
+    picker has its own window draw + visible loop wired — leaving ch4 enabled here corrupts the field BG3 layer
+    (Cecil walks on a split screen) since the engine's RTL goes back to the field via EventCmd_f7 without any
+    teardown.
+    """
+        rep #0x20
+        lda.w key_item_rolling_base_scroll
+        cmp.w #0xFFFF
+        bne _key_item_hdma_already_init
+        lda.l 0x7E019F
+        sta.w key_item_rolling_base_scroll
+        sep #0x20
+        jsr.w _key_item_init_hdma_channel
+        lda #KEY_ITEM_HDMA_CHANNEL_BIT
+        sta.l 0x7E1BAE
+        sta.w key_item_hdma_enable
+        rts
+
+    _key_item_hdma_already_init:
+        sep #0x20
+        rts
+
+
+    _key_item_init_hdma_channel:
+    """Configure HDMA ch4: DIRECT mode, dest BG3VOFS ($2112), source = picker shadow table at $7E:9900."""
+        php
+        sep #0x20
+        jsr.w update_key_item_scroll_hdma
+        lda #0x02
+        sta.l KEY_ITEM_HDMA4_CTRL
+        lda #0x12
+        sta.l KEY_ITEM_HDMA4_DEST
+        rep #0x20
+        lda.w #KEY_ITEM_HDMA_TABLE_ADDR
+        sta.l KEY_ITEM_HDMA4_SRC_LO
+        sep #0x20
+        lda #KEY_ITEM_HDMA_BANK
+        sta.l KEY_ITEM_HDMA4_SRC_BANK
+        plp
+        rts
+
+
+    key_item_render_item_to_slot:
+    """Render filtered item from $7E:0712 + edge_row*Item.__size into BG3 buffer at $7E:D600 + slot_index*128 + 0x44."""
+        php
+        phb
+        lda #0x7E
+        pha
+        plb
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x5a
+        pha
+        lda.b 0x29
+        pha
+        lda.b 0x45
+        pha
+        lda.b 0x33
+        pha
+        sep #0x20
+        lda.b 0x5d
+        pha
+        lda.b 0xDB
+        pha
+        rep #0x20
+        lda.w #0xD600
+        sta.b 0x29
+        sep #0x20
+        lda.w key_item_rolling_edge_row
+        asl
+        clc
+        adc #0x12
+        sta.b 0x5a
+        lda #0x07
+        adc #0x00
+        sta.b 0x5b
+        rep #0x20
+        lda.b 0x5a
+        tax
+        sep #0x20
+        lda.l 0x7E0000 + Item.id, x
+        pha
+        lda.l 0x7E0000 + Item.qty, x
+        sta.b 0x5C
+        stz.b 0x34
+        pla
+        jsr.l check_can_use_item_trampoline
+        lda.w key_item_rolling_slot_index
+        sta.b 0x5d
+        rep #0x20
+        lda.w key_item_rolling_slot_index
+        and.w #0x00FF
+        xba
+        lsr
+        clc
+        adc.w #0x0444
+        tay
+        sep #0x20
+        jsr.l draw_item_slot_inner_trampoline
+        pla
+        sta.b 0xDB
+        pla
+        sta.b 0x5d
+        rep #0x20
+        pla
+        sta.b 0x33
+        pla
+        sta.b 0x45
+        pla
+        sta.b 0x29
+        pla
+        sta.b 0x5a
+        rep #0x10
+        ply
+        plx
+        pla
+        plb
+        plp
+        rts
+
+
+    key_item_render_all:
+    """
+    Replacement for original UpdateItemText. Original just clears $0774 (text-buffer scratch) and walks $0712 to lay
+    out 4x4 grid into the BG3 buffer at $7E:D600. We replace with engine_init_rolling_buffer which renders 6
+    single-col slots from $0712 into the same buffer. Original NMI's BG3 transfer then pushes them to VRAM as part
+    of the existing item-window flow ($EB=$01 latched by original preamble at $00:AF53).
+    """
+        php
+        rep #0x10
+        sep #0x20
+        jsr.l key_item_init_impl
+        ; engine's ensure_hdma turned $1BAE bit 4 on; clear so original NMI
+        ; doesn't try to drive HDMA we haven't fully wired (per-scanline
+        ; bands not yet matched to the picker rows).
+        lda #0x00
+        sta.l 0x7E1BAE
+        lda #0x00
+        sta.l 0x00420C
+        plp
+        rtl
+
+
+    clear_key_item_slot:
+    """Blank one tilemap row at slot_index in the BG3 buffer."""
+        php
+        phb
+        lda #0x7E
+        pha
+        plb
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x29
+        pha
+        lda.w #0xD600
+        sta.b 0x29
+        lda.w key_item_rolling_slot_index
+        and.w #0x00FF
+        xba
+        lsr
+        clc
+        adc.w #0x0044
+        tay
+        sep #0x20
+        ldx.w #0x0000
+
+    _clear_key_loop:
+        lda #0x00
+        sta (0x29), y
+        iny
+        inx
+        cpx.w #0x0040
+        bne _clear_key_loop
+        rep #0x20
+        pla
+        sta.b 0x29
+        rep #0x10
+        ply
+        plx
+        pla
+        plb
+        plp
+        rts
+
+    key_item_init_filter:
+    """
+    Filter $1440 -> $0712. Faithful inline port of original InitItemList ($00:B2D5 in actual ROM, off-by-2 from
+    ff4decomp notes). Clears the 96-byte filter buffer, walks 48 inventory items, copies (id, qty) pairs whose IDs
+    are key items: [$CE..$E6] u [$EB..$FD].
+    """
+        php
+        phb
+        sep #0x20
+        rep #0x10
+        lda #0x7E
+        pha
+        plb
+        ldx.w #0x0000
+
+    _filter_clear:
+        stz.w 0x0712, x
+        inx
+        cpx.w #0x0060
+        bne _filter_clear
+        ldx.w #0x0000
+        ldy.w #0x0000
+
+    _filter_walk:
+        lda.w 0x1440, x
+        cmp #0xCE
+        bcc _filter_next
+        cmp #0xE7
+        bcc _filter_accept
+        cmp #0xEB
+        bcc _filter_next
+        cmp #0xFE
+        bcs _filter_next
+
+    _filter_accept:
+        sta.w 0x0712, y
+        lda.w 0x1441, x
+        sta.w 0x0713, y
+        iny
+        iny
+
+    _filter_next:
+        inx
+        inx
+        cpx.w #0x0060
+        bne _filter_walk
+        plb
+        plp
+        rts
+
+    update_key_item_scroll_hdma:
+    """Build the key-item HDMA shadow table via the shared engine."""
+        ; Build key-item picker HDMA scroll table. Inlined from the former
+        ; engine_update_scroll_hdma macro for the same reason as treasure.
+    {
+        php
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x40
+        pha
+        lda.b 0x42
+        pha
+        ldx.w #0x0000
+        jsr.w _key_item_hdma_header
+        stz.b 0x42
+
+    _row_loop:
+        lda.w key_item_rolling + RollingBufferState.buffer_pos
+        and.w #0x00FF
+        clc
+        adc.b 0x42
+
+    _mod_loop:
+        cmp.w #KEY_ITEM_BUFFER_SLOTS
+        bcc _mod_done
+        sec
+        sbc.w #KEY_ITEM_BUFFER_SLOTS
+        bra _mod_loop
+
+    _mod_done:
+        asl
+        asl
+        asl
+        asl
+        sta.b 0x40
+        lda.b 0x42
+        and.w #0x00FF
+        asl
+        asl
+        asl
+        asl
+        eor.w #0xFFFF
+        inc
+        clc
+        adc.b 0x40
+        clc
+        adc.w key_item_rolling + RollingBufferState.base_scroll
+        sta.b 0x40
+        sep #0x20
+        lda #16
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.b 0x40
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        inx
+        inx
+        rep #0x20
+        inc.b 0x42
+        lda.b 0x42
+        cmp.w #KEY_ITEM_VISIBLE_ITEMS
+        bcs _row_loop_done
+        jmp.w _row_loop
+
+    _row_loop_done:
+        jsr.w _key_item_hdma_footer
+        sep #0x20
+        lda #0x00
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        jsr.w _key_item_hdma_signal
+        rep #0x20
+        pla
+        sta.b 0x42
+        pla
+        sta.b 0x40
+        ply
+        plx
+        pla
+        plp
+        rts
+    }
+
+    _key_item_hdma_header:
+    """Picker HDMA header — 112 lines at BASE (top half = field map preserved)."""
+        sep #0x20
+        lda #112
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w key_item_rolling_base_scroll
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _key_item_hdma_footer:
+    """Picker HDMA footer — 16 lines at BASE+16 to hide prefetch slot."""
+        sep #0x20
+        lda #16
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w key_item_rolling_base_scroll
+        clc
+        adc.w #16
+        sta.l KEY_ITEM_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _key_item_hdma_signal:
+    """NMI shadow-copy signal — set both picker copy_pending + shared $1BB6 mirror."""
+        sep #0x20
+        lda #0x01
+        sta.w key_item_hdma_copy_pending
+        rts
+
+
+    key_item_init_impl:
+    """
+    Init key-item picker (filter $1440 -> $0712 then engine init). State
+    + hook far-ptrs live at $7E:9C60 (relocated out of $1B00-$1BFF).
+    """
+        jsr.w key_item_init_filter
+        php
+        rep #0x30
+        sep #0x20
+        lda.b #KEY_ITEM_BUFFER_SLOTS
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.visible_rows
+        lda.b #0x02
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.slot_height_tiles
+        ; item_list_ptr = $7E:0712 (filtered key-item array)
+        lda.b #0x12
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.item_list_ptr
+        lda.b #0x07
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.item_list_ptr + 1
+        lda.b #0x7E
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.item_list_ptr + 2
+        lda.b #KEY_ITEM_TOTAL_ITEMS
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.item_count
+        lda.b #0x04
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.hdma_channel
+        lda.b #0x80
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.vwf_cfg_ptr
+        lda.b #0x70
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.vwf_cfg_ptr + 1
+        lda.b #0x70
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.vwf_cfg_ptr + 2
+        lda.b #key_item_fn_render_slot_trampoline & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_render_slot
+        lda.b #( key_item_fn_render_slot_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_render_slot + 1
+        lda.b #( key_item_fn_render_slot_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_render_slot + 2
+        lda.b #key_item_fn_update_hdma_trampoline & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_update_hdma
+        lda.b #( key_item_fn_update_hdma_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_update_hdma + 1
+        lda.b #( key_item_fn_update_hdma_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_update_hdma + 2
+        lda.b #key_item_fn_draw_window_trampoline & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_draw_window
+        lda.b #( key_item_fn_draw_window_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_draw_window + 1
+        lda.b #( key_item_fn_draw_window_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.fn_draw_window + 2
+        lda.b #ROLLING_MENU_ID_KEY_ITEM
+        sta.l 0x7E0000 + key_item_rolling + RollingBufferState.menu_id
+        plp
+        php
+        rep #0x10
+        ldx.w #key_item_rolling
+        jsr.l rolling_engine.rolling_engine_init
+        plp
+        rtl
+
+    key_item_fn_render_slot_trampoline:
+    """Bank-20 RTL wrapper around `key_item_render_item_to_slot`."""
+        php
+        jsr.w key_item_render_item_to_slot
+        plp
+        rtl
+
+    key_item_fn_update_hdma_trampoline:
+    """Bank-20 RTL wrapper around `key_item_ensure_hdma_initialized`."""
+        php
+        jsr.w key_item_ensure_hdma_initialized
+        plp
+        rtl
+
+    key_item_fn_draw_window_trampoline:
+    """Bank-20 RTL wrapper around `_key_item_draw_window`."""
+        php
+        jsr.w _key_item_draw_window
+        plp
+        rtl
+
+
+    _key_item_draw_window:
+    """
+    Picker is invoked from inside original ShowItemWindow which already drew the
+    picker frame via its IRQ slide. No-op.
+    """
+        rts
+
+    key_item_start_scroll_down_impl:
+    """Picker: kick scroll-down state machine via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + key_item_scroll_pos
+        ldx.w #key_item_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_down
+        plp
+        rtl
+
+    key_item_start_scroll_up_impl:
+    """Picker: kick scroll-up state machine via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + key_item_scroll_pos
+        ldx.w #key_item_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_up
+        plp
+        rtl
+
+    key_item_update_scroll_frame_impl:
+    """Picker: per-frame scroll animation tick."""
+        php
+        rep #0x10
+        ldx.w #key_item_rolling
+        jsr.l rolling_engine.rolling_engine_update_scroll_frame
+        plp
+        rtl
+
+    key_item_finish_scroll_impl:
+    """Picker: end-of-animation cleanup via the bank-20 engine."""
+        php
+        rep #0x10
+        lda.l 0x7E0000 + key_item_scroll_pos
+        ldx.w #key_item_rolling
+        jsr.l rolling_engine.rolling_engine_finish_scroll
+        plp
+        rtl
+
+    key_item_refresh_slots_impl:
+    """Picker: re-render all slots via the bank-20 engine."""
+        php
+        sep #0x20
+        rep #0x10
+        lda.l 0x7E0000 + key_item_scroll_pos
+        ldx.w #key_item_rolling
+        jsr.l rolling_engine.rolling_engine_refresh_slots
+        plp
+        rtl
+}
diff --git a/src/ingame/key_item_picker_patches.s b/src/ingame/key_item_picker_patches.s
index dd9c243..608c8e0 100644
--- a/src/ingame/key_item_picker_patches.s
+++ b/src/ingame/key_item_picker_patches.s
@@ -31,6 +31,14 @@ Patched:
 
 
 .if TREASURE_INVENTORY_ROLLING {
+; TODO : wire `jsr.l key_item_render_all` at $00:AF4D once the picker
+; has a VRAM strategy that doesn't garble the room/map underneath. The
+; picker triggers from event scripts mid-map (EventCmd_f7), so unlike
+; the field menu we cannot blow away the BG3 CHR slice the room is
+; using. Two viable paths : (a) save/restore around picker entry/exit,
+; (b) reserve picker-only VRAM that the map provably never touches.
+; Engine code at `key_item_render_all` + `key_item_init_impl` is in
+; place but currently unreachable.
 ; UpdateItemText at $00:B22B reads scroll pos $BA, multiplies by 4
 ; (asl asl) for 2-col x 2-byte stride into $0712. Single-col layout
 ; needs stride = 2, so replace the 2nd `asl` at $00:B232 with NOP.
@@ -58,47 +66,51 @@ Patched:
     lda #0x04
 
 ; Replace the inline id*9 multiplier at $00:B253-B26C with a
-; jsl multiply_by_12 chain. A is item-id on entry (just loaded via
-; lda $0712,x at $B24B), returns A=id*12. Move into X for the
-; existing inner-loop indexed `lda.l assets_items_dat, x` read.
-; The original block was 26 bytes ($B253..$B26C); replacement uses
-; 9 bytes, padded with NOP to keep downstream instruction
-; addresses ($B26F lda#, $B273 lda.l ...) anchored.
+; jsl multiply_by_17 chain. A is item-id on entry (just loaded via
+; lda $0712,x at $B24B), returns A = id * ITEM_UNLEASHED_RECORD_SIZE.
+; Move into X for the existing inner-loop indexed
+; `lda.l assets_items_unleashed_dat, x` read. The original block was
+; 26 bytes ($B253..$B26C); replacement uses 9 bytes, padded with NOP
+; to keep downstream instruction addresses ($B26F lda#, $B273 lda.l
+; ...) anchored.
     *=0x00B253
     rep #0x10
-    jsr.l multiply_by_12
+    jsr.l multiply_by_17
     tax
     inx
     pad_nop(18)
 
 ; Inner-name-write loop count at $00:B26F: original `lda #$08`
-; (8 letters per name). Bump to ITEM_NAME_TEXT_SIZE.
+; (8 letters per name). Bump to ITEM_UNLEASHED_TEXT_SIZE.
     *=0x00B26F
-    lda #ITEM_NAME_TEXT_SIZE
+    lda #ITEM_UNLEASHED_TEXT_SIZE
 
 ; Item-name table source at $00:B273: original `lda.l $0F8000,x`
-; (JP layout). Redirect to French assets_items_dat.
+; (JP layout). Redirect to assets_items_unleashed_dat.
     *=0x00B273
-    lda.l assets_items_dat, x
+    lda.l assets_items_unleashed_dat, x
 
 ; Make both column-toggle branches advance Y by 24 (full text-buffer
 ; row, 12 chars). Both → `adc #$18` so every item lands on its own
-; row regardless of which column the toggle picks.
+; row regardless of which column the toggle picks. Keep at 24
+; for now ; widening to fit a 16-char name + colon + qty requires
+; revisiting the picker's tilemap row stride too.
     *=0x00B2B5
     adc #0x18
     *=0x00B2BF
     adc #0x18
 
 ; Item-text layout post-name: original writes ":" at +8 ($077C), tens
-; at +9 ($077D), ones at +10 ($077E). Names were 8 chars (slots 0..7).
-; French names are 11 chars (slots 0..10) + 1 tile spacer = 12, so
-; push the qty trio out by +12 (one full tile beyond the name):
+; at +9 ($077D), ones at +10 ($077E). 12-char names pushed the trio
+; to $0780/81/82. 16-char names need another +4 to land past the name
+; (assets_items_unleashed_dat = symbol + 16 chars), so trio sits at
+; $0784/85/86 with the same 1-tile spacer between name and colon.
     *=0x00B28E
-    sta 0x0780, y
+    sta 0x0784, y
     *=0x00B2A4
-    sta 0x0781, y
+    sta 0x0785, y
     *=0x00B2A9
-    sta 0x0782, y
+    sta 0x0786, y
 
 ; Single-col picker has no col-1 to move cursor to. NOP the JOY_RIGHT
 ; check at $00:AFB0 by replacing the AND mask with $00 — beq always
diff --git a/src/ingame/main.s b/src/ingame/main.s
index f401ccc..efcf29b 100644
--- a/src/ingame/main.s
+++ b/src/ingame/main.s
@@ -219,11 +219,22 @@ end:
 *=0x1b349
     lda.w assets_magic_dat, y
 
-; Copy of save and restore vram routines from menu, save 0x1300 instead of 0x1000 and store it to the sram
+; Save / restore covers VRAM byte $4000-$5FFF ($2000 bytes) instead of
+; vanilla's $1000. The extra $1000 bytes reach past the static-font /
+
+; vanilla BG3 CHR area to cover the full VWF region union :
+;   Region 1   $5000-$56A0  (field / treasure item-name CHR)
+;   Region 1B  $56E0-$5AA0  (drops item-name CHR, treasure popup)
+;   Region D   $5800-$5FF0  (item description CHR)
+; Without the extension, menu glyph bytes at $5300+ stayed in VRAM
+; forever (overworld never writes those addresses), so any BG tilemap
+; entry referencing VWF tile_ids $100+ rendered the leftover glyphs
+; over the overworld map. SRAM buffer at $70:5000-$70:6FFF stays
+; clear of the battle-magic region at $70:7000.
 {
     *=0x14ff62
     sram_buffer = 0x705000
-    save_size = 0x1300
+    save_size = 0x2000
     phb
     tdc
     pha
diff --git a/src/ingame/places_names_window.s b/src/ingame/places_names_window.s
index 74f57d3..cafe514 100644
--- a/src/ingame/places_names_window.s
+++ b/src/ingame/places_names_window.s
@@ -1,9 +1,14 @@
 """Tilemap rows for the top + bottom edges of the location-name window."""
 
-places_top_window:
-"""Tilemap row for the top edge of the location-name window (60 bytes, 30 tile/attr pairs)."""
-    .db 0x16, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x18, 0x20  ; noqa: E501
 
-places_bottom_window:
-"""Tilemap row for the bottom edge of the location-name window (60 bytes, 30 tile/attr pairs)."""
-    .db 0x1B, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1D, 0x20  ; noqa: E501
+.include "../bank20.i"
+
+.alloc places_names_window_block in bank20_reloc {
+    places_top_window:
+    """Tilemap row for the top edge of the location-name window (60 bytes, 30 tile/attr pairs)."""
+        .db 0x16, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x17, 0x20, 0x18, 0x20  ; noqa: E501
+
+    places_bottom_window:
+    """Tilemap row for the bottom edge of the location-name window (60 bytes, 30 tile/attr pairs)."""
+        .db 0x1B, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1C, 0x20, 0x1D, 0x20  ; noqa: E501
+}
diff --git a/src/ingame/shop.s b/src/ingame/shop.s
index aae0818..6c00557 100644
--- a/src/ingame/shop.s
+++ b/src/ingame/shop.s
@@ -47,18 +47,38 @@ small-VWF item descriptions.
 
 *=0x01C350
     load_system_menu_text_pointer(shops.welcome_and_actions)
+    ; Route the owner greeting ("Puis-je vous aider ?") through the small-VWF
+    ; description region ; the slimmed welcome_and_actions block holds only the
+    ; "Achat Vente Sortir" action labels rendered by the vanilla engine.
+
+*=0x01C353
+    jmp.w shop_welcome_text_hook
 
 *=0x01C43F
     load_system_menu_text_pointer(shops.quantity)
+    ; Route the buy-flow welcome ("Que désirez vous ?") through the small-VWF
+    ; description region ; vanilla quantity block (Quantité + "1") continues to
+    ; render via $8301 inside the hook.
+
+*=0x01C442
+    jsr.w shop_quantity_text_hook
 
 *=0x01c7e4
     load_system_menu_text_pointer(shops.quantity)
 
+*=0x01C7E7
+    jsr.w shop_quantity_text_hook
+
 *=0x01C568
     load_system_menu_text_pointer(shops.gils + 2)
 
 *=0x01c74e
     load_system_menu_text_pointer(shops.thank_you_window)
+    ; Route the owner thank-you ("Merci !") through the small-VWF description
+    ; region ; the window itself still draws via vanilla $82FB inside the hook.
+
+*=0x01C751
+    jsr.w shop_thanks_text_hook
 
 *=0x01c962
     load_system_menu_text_pointer(shops.sell_window)
diff --git a/src/ingame/treasure_rolling.s b/src/ingame/treasure_rolling.s
index 5e7f4d2..984b23d 100644
--- a/src/ingame/treasure_rolling.s
+++ b/src/ingame/treasure_rolling.s
@@ -40,7 +40,12 @@ TREASURE_ITEM_LIST_HEIGHT := 80  ; 5 items × 16 pixels
 ; struct layout as the field-menu state block — fields resolved via
 ; `RollingBufferState` from src/items.i so any layout change applies
 ; uniformly across profiles.
-treasure_rolling := 0x1BD0
+; Treasure rolling state moved out of $1B00-$1BFF (which vanilla menu /
+; sprite code writes to via indexed STA past $1BEB) to a clean $7E:9C00
+; region. Engine path needs the full 35-byte struct (state + config +
+; hook far-ptrs) ; the macro path only ever touched the first 12 bytes
+; so the original $1BD0 base worked despite vanilla's later collisions.
+treasure_rolling := 0x9C00
 treasure_rolling_top_row := treasure_rolling + RollingBufferState.top_row
 treasure_rolling_buffer_pos := treasure_rolling + RollingBufferState.buffer_pos
 treasure_rolling_edge_row := treasure_rolling + RollingBufferState.edge_row
@@ -53,6 +58,18 @@ treasure_scroll_direction := treasure_rolling + RollingBufferState.scroll_direct
 treasure_transfer_pending := treasure_rolling + RollingBufferState.transfer_pending
 treasure_scroll_anim_offset := treasure_rolling + RollingBufferState.scroll_anim_offset
 treasure_hdma_copy_pending := treasure_rolling + RollingBufferState.hdma_copy_pending
+; Cooldown counter sitting one byte past the shared struct so the
+; engine layout stays untouched. Decremented each frame in
+; treasure_scroll_state_check ; non-zero means treasure_scroll_*_trigger
+; aborts and undoes vanilla's $1BB7 increment, debouncing the
+; "hold-DOWN auto-repeat fires every 2 frames" issue that scrolled the
+; inventory two items per visible tap.
+; RollingBufferState ends at offset 35 inclusive (menu_id byte added
+; in the engine port). Bump cooldown past that ; was +35 = collided
+; with menu_id and treasure_ensure_hdma_initialized's STZ wiped it,
+; sending engine dispatch to the wrong menu's HDMA path.
+treasure_scroll_cooldown := treasure_rolling + 36
+TREASURE_SCROLL_COOLDOWN_FRAMES := 0x18  ; ~24 frames between scrolls (long enough to outlast a typical button hold)
 
 ; Scroll State Constants
 TREASURE_SCROLL_STATE_IDLE := 0
@@ -108,883 +125,1100 @@ TREASURE_HDMA6_SRC_BANK := 0x4364
 TREASURE_HDMA6_IND_BANK := 0x4367
 HDMAEN := 0x420C  ; HDMA enable register
 
-init_treasure_inventory_hdma:
-"""
-Sets up HDMA channel 5 for per-scanline BG1 vertical scroll control
-Called when entering the item menu
-
-HDMA mode: $02 = write 2 bytes to same register, DIRECT mode (like FF6)
-Register: $210E (BG1VOFS)
-Table format: count_byte, lo_byte, hi_byte per entry, $00 to end
-"""
-
-
-    php
-    sep #0x20  ; 8-bit A
-
-; Build the HDMA data table in WRAM
-; Use UpdateMenuScrollHDMA directly - InitMenuHDMATable is redundant
-; since UpdateMenuScrollHDMA is called immediately after anyway
-    jsr.w update_treasure_scroll_hdma
-
-; Configure HDMA channel 6 for DIRECT mode
-; Must use long addressing - DB may be $7E but registers are at $00:43xx
-    lda #0x02  ; Mode: DIRECT, write 2 bytes to same PPU reg
-    sta.l TREASURE_HDMA6_CTRL  ; $004360
-
-    lda #0x12  ; BG3VOFS register ($2112) — treasure inventory is on BG3
-    sta.l TREASURE_HDMA6_DEST  ; $004361
-
-; Source = HDMA table in WRAM at $7E9800
-    rep #0x20  ; 16-bit A
-    lda.w #TREASURE_HDMA_TABLE_ADDR  ; $9800
-    sta.l TREASURE_HDMA6_SRC_LO  ; $004362-$004363
-    sep #0x20  ; 8-bit A
-    lda #TREASURE_HDMA_BANK  ; $7E
-    sta.l TREASURE_HDMA6_SRC_BANK  ; $004364
-
-; HDMA channel 5 is now enabled via shadow variable (treasure_hdma_enable)
-; The NMI hook at $8083 reads the shadow and writes to HDMAEN
-
-    plp
-    rts
-
-disable_treasure_inventory_hdma:
-"""
-Disables HDMA channel 5 when leaving item menu
-The shadow variable is cleared by menu_exit_hook
-"""
-    php
-    sep #0x20  ; 8-bit A
-    ; Shadow variable cleared by caller, NMI will write 0 to HDMAEN
-    plp
-    rts
-
-init_treasure_hdma_table:
-"""
-Builds direct mode HDMA table in WRAM at $7E9800
-Format: count, lo, hi per entry, $00 to end
-Initial state: all rows at BASE scroll (no circular buffer offset)
-"""
-    rep #0x30  ; 16-bit A, X, Y
-    php
-    pha  ; Save A
-    phx  ; Save X
-    phy  ; Save Y
-
-; Save DP bytes we'll use as scratch
-    lda.b 0x40
-    pha  ; Save $40-$41
-
-; X = table write offset
-    ldx.w #0x0000
-
-; Entry 0: Border area - 48 scanlines at BASE scroll
-    sep #0x20  ; 8-bit A for count byte
-    lda #48  ; 48 scanlines
-    sta.l TREASURE_HDMA_TABLE, x
-    inx
-    rep #0x20  ; 16-bit A for value
-    lda.w treasure_rolling_base_scroll
-    sta.l TREASURE_HDMA_TABLE, x
-    inx
-    inx
-
-; Entries 1-10: Item rows - 16 scanlines each at BASE scroll (initial)
-; For init, all rows use BASE (no circular buffer offset yet)
-    lda.w #0x0000
-    sta.b 0x40  ; Row counter
-
-_t_init_item_rows:
-    sep #0x20  ; 8-bit A for count
-    lda #16  ; 16 scanlines per item row
-    sta.l TREASURE_HDMA_TABLE, x
-    inx
-    rep #0x20  ; 16-bit A for value
-    lda.w treasure_rolling_base_scroll
-    sta.l TREASURE_HDMA_TABLE, x
-    inx
-    inx
-
-    inc.b 0x40
-    lda.b 0x40
-
-    cmp.w #1  ; 10 rows
-    bcc _t_init_item_rows
-
-; Entry 11: Below items - 16 scanlines at BASE + 16
-; Lock to show the bottom window border area
-    sep #0x20
-    lda #16  ; 16 scanlines
-    sta.l TREASURE_HDMA_TABLE, x
-    inx
-    rep #0x20
-    lda.w treasure_rolling_base_scroll
-    clc
-    adc.w #16  ; Lock at base + 16
-    sta.l TREASURE_HDMA_TABLE, x
-    inx
-    inx
-
-; End marker
-    sep #0x20
-    lda #0x00
-    sta.l TREASURE_HDMA_TABLE, x
-
-; Restore DP bytes
-    rep #0x20
-    pla
-    sta.b 0x40  ; Restore $40-$41
-
-; Restore registers
-    ply
-    plx
-    pla
-    plp
-    rts
-
-update_treasure_scroll_hdma:
-"""Build the treasure-menu HDMA scroll table via the shared engine."""
-    engine_update_scroll_hdma(treasure_rolling, TREASURE_HDMA_SHADOW, TREASURE_BUFFER_SLOTS, TREASURE_VISIBLE_ITEMS, _treasure_hdma_header, _treasure_hdma_footer, _treasure_hdma_signal)  ; noqa: E501
-
-_treasure_hdma_header:
-"""Treasure profile header: drops band (120 lines at BASE-16) + inventory border (8 lines at BASE)."""
-
-
-    sep #0x20
-    lda #120
-    sta.l TREASURE_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w treasure_rolling_base_scroll
-    sec
-    sbc.w #16
-    sta.l TREASURE_HDMA_SHADOW, x
-    inx
-    inx
-    sep #0x20
-    lda #8
-    sta.l TREASURE_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w treasure_rolling_base_scroll
-    sta.l TREASURE_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_treasure_hdma_footer:
-"""Treasure profile footer: 16 scanlines at BASE+16 (hides prefetch slot)."""
-
-
-    sep #0x20
-    lda #16
-    sta.l TREASURE_HDMA_SHADOW, x
-    inx
-    rep #0x20
-    lda.w treasure_rolling_base_scroll
-    clc
-    adc.w #16
-    sta.l TREASURE_HDMA_SHADOW, x
-    inx
-    inx
-    rts
-
-_treasure_hdma_signal:
-"""Treasure profile signal: set both treasure flag and shared $1BB6 mirror."""
-    sep #0x20
-    lda #0x01
-    sta.w treasure_hdma_copy_pending
-    sta.w treasure_hdma_copy_pending_shared
-    rts
-
-; Re-render all BUFFER_SLOTS (6) entries from $1440 using current
-; $1BB7 (scroll_pos) and the existing buffer_pos rotation. Called from
-; the redraw helper at $01:D933 after a swap completes — the swap
-; mutates $1440 in place and original's DrawInventoryList would redraw
-; the whole 48-item list, but we only need to refresh the 5 visible
-; slots + 1 prefetch. Crucially does NOT touch buffer_pos or any of
-; the state-machine bytes, so the cursor/scroll position the user
-; was on before the swap survives.
-
-treasure_refresh_slots_impl:
-"""Treasure profile: re-render all 6 slots without resetting scroll state (original redraw helper at $01:D933)."""
-    engine_refresh_slots(treasure_rolling, 0x1BB7, TREASURE_BUFFER_SLOTS, _treasure_render_item_to_slot)
-
-init_treasure_rolling_buffer_impl:
-"""Init treasure rolling buffer (5 visible + prefetch slot 6)."""
-    engine_init_rolling_buffer(treasure_rolling, TREASURE_BUFFER_SLOTS, _treasure_draw_inventory_window, treasure_ensure_hdma_initialized, _treasure_render_item_to_slot)  ; noqa: E501
-
-_treasure_draw_inventory_window:
-"""Treasure menu draws a 5-row inventory window so the bottom border lands on the rolling-buffer footer scanlines."""
-    rep #0x10
-    ldy.w #treasure_inventory_window
-    jsr.l draw_window_trampoline
-    sep #0x10
-    rts
-
-treasure_scroll_down_prepare:
-"""Treasure profile scroll-down pre-render."""
-    engine_scroll_down_prepare(treasure_rolling, 0x1BB7, TREASURE_SCROLL_LIMIT, TREASURE_VISIBLE_ITEMS, TREASURE_BUFFER_SLOTS, treasure_ensure_hdma_initialized, _treasure_render_item_to_slot, update_treasure_scroll_hdma)  ; noqa: E501
-
-treasure_scroll_up_prepare:
-"""Treasure profile scroll-up pre-render."""
-    engine_scroll_up_prepare(treasure_rolling, 0x1BB7, TREASURE_BUFFER_SLOTS, treasure_ensure_hdma_initialized, _treasure_render_item_to_slot, update_treasure_scroll_hdma)  ; noqa: E501
-
-; _treasure_render_item_to_slot
-; Renders an item to a specific circular buffer slot in the tilemap.
-;
-; Input: treasure_rolling_edge_row = item index (0-47) for data lookup
-;        treasure_rolling_slot_index = slot index (0-11) for destination
-;
-; Strategy: Set up $5d = slot_index (for Y position calculation),
-;           $5a = pointer to item data, then call game's DrawItemSlot.
-
-_treasure_render_item_to_slot:
-    php
-    phb
-
-; Set data bank to $7E for WRAM access
-    lda #0x7E
-    pha
-    plb
-
-; Set 16-bit A and X/Y for consistent register handling
-    rep #0x30  ; 16-bit A and X/Y
-
-; Save registers and key direct page variables
-    pha
-    phx  ; Save X (16-bit)
-    phy  ; Save Y (16-bit)
-    lda.b 0x5a
-    pha
-    lda.b 0x29  ; Save tilemap buffer pointer
-    pha
-    lda.b 0x45  ; Save $45-$46 (used by game routines)
-    pha
-    lda.b 0x33  ; Save $33-$34 (used for tile attributes)
-    pha
-    sep #0x20  ; 8-bit A
-    lda.b 0x5d
-    pha
-    lda.b 0xDB  ; Save tile attribute byte
-    pha
-
-; Set $29 = $B600 for BG1 tilemap buffer
-    rep #0x20  ; 16-bit A (X/Y still 16-bit)
-    lda.w #0xD600  ; BG3 screen buffer (treasure inventory lives on BG3)
-    sta.b 0x29
-    sep #0x20  ; 8-bit A
-
-; Calculate item data pointer: $1440 + (edge_row * Item.__size)
-    lda.w treasure_rolling_edge_row
-    asl  ; * Item.__size (2 bytes per Item)
-    clc
-    adc #0x40  ; Low byte of $1440
-    sta.b 0x5a
-    lda #0x14  ; High byte of $1440
-    adc #0x00  ; Add carry
-    sta.b 0x5b
-
-; Load Item.id and Item.qty from ($5A) via long-addressing into WRAM.
-    rep #0x20
-    lda.b 0x5a  ; Pointer value = $1440 + edge_row * Item.__size
-    tax
-    sep #0x20
-    lda.l 0x7E0000 + Item.id, x
-    pha  ; Save Item.id for CheckCanUseItem
-    lda.l 0x7E0000 + Item.qty, x
-    sta.b 0x5C  ; Store qty in $5C
-
-; Call CheckCanUseItem to set palette in $DB
-; Input: A = item ID
-; Output: $DB = $00 (usable) or $04 (not usable)
-    stz.b 0x34  ; Clear $34 (no priority/flip bits)
-    pla  ; Restore item ID
-    jsr.l check_can_use_item_trampoline  ; bank-$01 trampoline for original @ $A25D (sets $DB)
-
-; Set $5d = slot_index (for AND #$01 check, but we patched to AND #$00)
-    lda.w treasure_rolling_slot_index
-    sta.b 0x5d
-
-; Calculate Y = slot_index * 128 + $44
-; Y is the tilemap offset for this slot
-; +64 = 1 BG row (top window border at staging row 0)
-; +4 for left margin (2 tiles)
-    rep #0x20  ; 16-bit A (X/Y already 16-bit)
-    lda.w treasure_rolling_slot_index
-    and.w #0x00FF  ; Clear high byte
-    xba  ; Swap bytes: A = slot * 256
-    lsr  ; A = slot * 128
-    clc
-    adc.w #0x0044  ; + 68 (64 border + 4 margin)
-    tay  ; Y = tilemap offset (16-bit transfer)
-    sep #0x20  ; 8-bit A
-
-; Check for trash can item ($FF) - needs special 2x2 tile graphic
-    lda (0x5a)  ; Load item ID
-    cmp #0xFF
-    bne _t_not_trash_item
-    jsr.w draw_trash_treasure  ; Draw trash icon
-    bra _t_skip_draw_item_slot
-
-_t_not_trash_item:
-    ; Clear the 2x2 trash can area first (in case we scrolled from trash position)
-    jsr.w _clear_treasure_trash_area
-
-; Call DrawItemSlot inner at $A1ED
-; Expects: Y = tilemap offset, ($5A) = item pointer, ($29) = tilemap base
-    jsr.l draw_item_slot_inner_trampoline  ; bank-$01 trampoline for original @ $A1ED
-
-_t_skip_draw_item_slot:
-
-; Restore direct page variables and registers (reverse order)
-    pla
-    sta.b 0xDB  ; Restore tile attribute byte
-    pla
-    sta.b 0x5d
-    rep #0x20  ; 16-bit A
-    pla
-    sta.b 0x33  ; Restore $33-$34 (tile attributes)
-    pla
-    sta.b 0x45  ; Restore $45-$46 (used by game routines)
-    pla
-    sta.b 0x29  ; Restore tilemap buffer pointer
-    pla
-    sta.b 0x5a
-    rep #0x10  ; 16-bit X/Y for pop (match push)
-    ply  ; Restore Y (16-bit)
-    plx  ; Restore X (16-bit)
-    pla  ; Restore A (16-bit - still in 16-bit A from above)
-
-    plb
-    plp
-    rts
-
-treasure_ensure_hdma_initialized:
-"""
-Lazy initialization: captures $93 and sets up HDMA on first scroll.
-Called from scroll prepare functions.
-Checks if base_scroll == 0xFFFF (sentinel) and if so, initializes.
-"""
-
-; Check if already initialized (base_scroll != 0xFFFF)
-    rep #0x20  ; 16-bit A
-    lda.w treasure_rolling_base_scroll
-    cmp.w #0xFFFF
-    bne _t_hdma_already_init
-
-; Capture original BG3VOFS shadow ($9F). Original treasure menu uses
-; $9F = -120 to position items at screen scanline 120; we mirror that.
-    lda.l 0x7E019F
-    sta.w treasure_rolling_base_scroll
-
-; Initialize HDMA channel configuration
-    sep #0x20  ; Back to 8-bit for InitMenuInventoryHDMA
-    jsr.w init_treasure_inventory_hdma
-
-; Enable HDMA via the SHARED field-menu shadow at $1BAE. The existing
-; field NMI hook (`field_menu_nmi_dma_transfer_check_impl`) reads this
-; byte and copies the HDMA shadow→active table each frame.
-; Treasure-only `treasure_hdma_enable` ($1BD6) is kept as a tracking
-; flag but isn't read by the NMI path.
-; OR-in ch6 enable bit ($40) so drops's ch4 bit ($10) — set by
-; drops_ensure_hdma_initialized at menu open — survives. Plain
-; `sta` would clobber the drops enable; same lazy-init order issue
-; as treasure_force_hdma_setup which uses $F9 (= ch7|ch6|ch5|ch4|ch3|ch0).
-    lda.l 0x7E1BAE
-    ora #0x40  ; Channel 6 enable (treasure rolling buffer)
-    sta.l 0x7E1BAE
-    sta.w treasure_hdma_enable
-    rts
-
-_t_hdma_already_init:
-    sep #0x20  ; Restore 8-bit mode
-    rts
-
-; STATE MACHINE ROUTINES (FF6-style non-blocking scroll)
-
-treasure_scroll_state_check:
-"""
-Called at main loop entry ($019FF2) to handle scroll animation frames.
-If scrolling is active, processes one frame and skips input handling.
-
-Returns: Carry clear = process input normally
- Carry set = skip input (still scrolling)
-"""
-
-
-    php
-    sep #0x20  ; 8-bit A
-
-; Check if we're scrolling
-    lda.w treasure_scroll_state
-    beq _t_scroll_state_idle
-
-; We're scrolling - process one animation frame
-    jsr.l treasure_update_scroll_frame_impl
-
-; Check if scroll finished
-    lda.w treasure_scroll_remaining
-    bne _t_scroll_still_active
-
-; Scroll finished - clean up and return to idle
-    jsr.l treasure_finish_scroll_impl
-
-_t_scroll_state_idle:
-    plp
-    clc  ; Carry clear = process input
-    rts
-
-_t_scroll_still_active:
-    plp
-    sec  ; Carry set = skip input
-    rts
-
-treasure_start_scroll_down_impl:
-"""Treasure profile: kick scroll-down state machine."""
-    engine_start_scroll_down(treasure_rolling, 0x1BB7, TREASURE_VISIBLE_ITEMS, TREASURE_BUFFER_SLOTS, TREASURE_SCROLL_TOTAL_PIXELS, TREASURE_SCROLL_PIXELS_PER_FRAME, treasure_ensure_hdma_initialized, _treasure_render_item_to_slot, update_treasure_scroll_hdma)  ; noqa: E501
-
-treasure_start_scroll_up_impl:
-"""Treasure profile: kick scroll-up state machine."""
-    engine_start_scroll_up(treasure_rolling, 0x1BB7, TREASURE_BUFFER_SLOTS, TREASURE_SCROLL_TOTAL_PIXELS, treasure_ensure_hdma_initialized, _treasure_render_item_to_slot, update_treasure_scroll_hdma)  ; noqa: E501
-
-treasure_update_scroll_frame_impl:
-"""Treasure profile: per-frame scroll animation tick."""
-    engine_update_scroll_frame(treasure_rolling, TREASURE_SCROLL_PIXELS_PER_FRAME, update_treasure_scroll_hdma)
-
-treasure_finish_scroll_impl:
-"""Treasure profile: end-of-animation pre-render + cleanup."""
-    engine_finish_scroll(treasure_rolling, 0x1BB7, TREASURE_VISIBLE_ITEMS, TREASURE_BUFFER_SLOTS, TREASURE_TOTAL_ITEMS, _treasure_render_item_to_slot, update_treasure_scroll_hdma)  ; noqa: E501
-
-; Inventory Rolling Buffer Patches - Relocated to Bank $20
-; These routines are called via JSL from bank $01 hooks.
-
-.if INVENTORY_ROLLING_BUFFER {
-treasure_menu_entry_hook_impl:
-"""Treasure profile: menu-entry implementation (HDMA capture + flush)."""
-    stz.w 0x1B1F
-    lda #0x00
-    sta.l 0x7E1BAE
-; treasure_hdma_enable
-    stz.w treasure_scroll_state
-    stz.w treasure_scroll_remaining
-    stz.w treasure_scroll_direction
-    stz.w treasure_transfer_pending
-    stz.w treasure_hdma_copy_pending
-; Clear HDMA copy flag
-    stz.w treasure_scroll_anim_offset
-; Clear low byte
-    stz.w treasure_scroll_anim_offset + 1
-; Clear high byte
-; Initialize cursor column to 0 for single-column mode
-; This ensures $1b22 is always 0 even if it had a value from previous menu
-    stz.w 0x1B22
-; cursor_x = 0
-; InitMenuRollingBuffer_Impl is called later via patched JSR at $9F7B
-    rtl
-
-treasure_menu_exit_hook_impl:
-"""Treasure menu teardown: zero HDMA shadow, full 12-byte state, ch6 regs, original $1BC6 flag."""
-    php
-    sep #0x20
-    lda #0x00
-    sta.l 0x7E1BAE
-; treasure_hdma_enable shadow off
-    sta.l 0x7E1BC6
-; restore original "in treasure menu" flag (was original `stz $1BC6` at $01:D7E6 before the hook patch)
-    sta.l 0x004360
-    sta.l 0x004361
-    sta.l 0x004362
-    sta.l 0x004363
-    sta.l 0x004364
-; HDMA6 ctrl/dest/src cleared
-    rep #0x20
-    lda.w #0x0000
-    sta.w treasure_rolling
-    sta.w treasure_rolling + 2
-    sta.w treasure_rolling + 4
-    sta.w treasure_rolling + 6
-    sta.w treasure_rolling + 8
-    sta.w treasure_rolling + 10
-    plp
-    jsr.l reset_sprites_trampoline
-    rtl
-; treasure_swap_redraw_hook_impl_body
-; Called after item swap to redraw visible items correctly.
-; Must render to the correct circular buffer slots based on current buffer_pos.
-; Does NOT reset buffer_pos - we stay at the current scroll position.
-
-treasure_swap_redraw_hook_impl_body:
-"""Treasure profile: post-swap re-render of all 6 slots."""
-    engine_swap_redraw(treasure_rolling, 0x1BB7, TREASURE_BUFFER_SLOTS, TREASURE_TOTAL_ITEMS, treasure_ensure_hdma_initialized, _treasure_render_item_to_slot, _clear_treasure_slot, update_treasure_scroll_hdma)  ; noqa: E501
-
-
-; _clear_treasure_slot
-; Clears a single inventory slot in the tilemap buffer.
-; Input: treasure_rolling_slot_index = slot to clear (0-10)
-; Used when item index is out of bounds (>= 48)
-
-_clear_treasure_slot:
-    php
-    phb
-; Set data bank to $7E for WRAM access
-    lda #0x7E
-    pha
-    plb
-; Save registers - use 16-bit mode for consistent push/pop
-    rep #0x30
-; 16-bit A and X/Y
-    pha
-    phx
-    phy
-    lda.b 0x29
-    pha
-; Set $29 = $B600 for BG1 tilemap buffer
-    lda.w #0xD600  ; BG3 screen buffer (treasure inventory lives on BG3)
-    sta.b 0x29
-; Calculate Y = slot_index * 128 + 70
-    lda.w treasure_rolling_slot_index
-    and.w #0x00FF
-    xba
-; A = slot * 256
-    lsr
-; A = slot * 128
-    clc
-    adc.w #0x0044
-; + 68 (64 border + 4 margin)
-    tay
-; Y = tilemap offset (16-bit)
-    sep #0x20
-; 8-bit A for tile writes (X/Y stay 16-bit)
-; Clear the item name area (12 tiles = 24 bytes)
-; Use $FF as blank tile
-; Note: X is 16-bit but we only use low byte; loop works correctly
-    ldx.w #12
-; 12 tiles for item name + quantity (force 16-bit immediate)
-
-_t_clear_slot_loop:
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-    iny
-    lda #0x04
-; Palette 4 (matches normal items)
-    sta (0x29), y
-    iny
-    dex
-    bne _t_clear_slot_loop
-; Restore registers - must match push mode
-    rep #0x20
-; 16-bit A for pop (X/Y already 16-bit)
-    pla
-    sta.b 0x29
-    ply
-    plx
-    pla
-    plb
-    plp
-    rts
-; _clear_treasure_trash_area
-; Clears the 2x2 trash can area with blank tiles ($FF).
-; Called before drawing normal items to remove any leftover trash icon.
-; Input: Y = tilemap offset
-;        ($29) = tilemap base
-;
-
-_clear_treasure_trash_area:
-    phy
-; Save Y
-; First row: 2 tiles
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-; Second row: Y + 64 from start
-    ply
-; Restore original Y
-    phy
-; Save again
-    rep #0x20
-    tya
-    clc
-    adc.w #64
-    tay
-    sep #0x20
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-    iny
-    lda #0x00
-    sta (0x29), y
-    ply
-; Restore Y
-    rts
-
-draw_trash_treasure:
-
-
-"""
-Draws the trash can 2x2 tile graphic for single-column inventory.
-Input: Y = tilemap offset (from slot calculation)
-($29) = tilemap base ($B600)
-treasure_rolling_slot_index = current slot
-Tiles: $04 (top-left), $05 (top-right), $06 (bottom-left), $07 (bottom-right)
-
-Tilemap format: [tile_number, attributes] pairs
-Each row is 64 bytes (32 tiles × 2 bytes)
-"""
-
-
-; Y points to start of item slot area
-; Draw 2x2 trash can icon, then clear remaining 10 tiles per row
-; Save starting Y for second row calculation
-    rep #0x20
-; 16-bit for push
-    phy
-; Save starting Y
-    sep #0x20
-; 8-bit A
-; First row: tiles $04, $05
-    lda #0x04
-; Tile $04 (top-left of trash can)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute: palette 0, no flip
-    sta (0x29), y
-    iny
-    lda #0x05
-; Tile $05 (top-right)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-; Clear remaining 13 tiles on first row (10 name + colon + 2 digits)
-    ldx.w #13
-
-_t_clear_row1:
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-    dex
-    bne _t_clear_row1
-; Restore starting Y and add 64 for second row
-    rep #0x20
-; 16-bit A
-    pla
-; Get starting Y
-    clc
-    adc.w #64
-; +64 bytes = next tilemap row
-    tay
-    sep #0x20
-; 8-bit A
-; Second row: tiles $06, $07
-    lda #0x06
-; Tile $06 (bottom-left)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-    lda #0x07
-; Tile $07 (bottom-right)
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-; Clear remaining 13 tiles on second row (10 name + colon + 2 digits)
-    ldx.w #13
-
-_t_clear_row2:
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-    iny
-    lda #0x00
-; Attribute
-    sta (0x29), y
-    iny
-    dex
-    bne _t_clear_row2
-    rts
-; NmiDmaTransferCheck_Impl moved to bank $20 (battle/inventory_rolling.s)
-; as field_menu_nmi_dma_transfer_check_impl to save space in bank $01
-; treasure_check_and_clear_count_impl
-; Called from DrawItemSlot to check item ID and handle count display.
-; - If item ID is 0: writes $FF tiles to clear count, skips to RTS
-; - If item ID is $FE: skips to RTS (no clearing needed)
-; - Otherwise: returns normally to draw count
-;
-; Input: $5a = pointer to item data, Y = tilemap offset, $29 = tilemap ptr
-; Modifies: A, return address on stack if skipping
-
-treasure_check_and_clear_count_impl:
-"""Treasure profile: drop the count column for empty/used slots (impl)."""
-    lda (0x5a)
-; Load item ID
-    beq _t_clear_count
-; If 0, clear and skip
-    cmp #0xFE
-; Check for special item $FE
-    bne _t_normal_return
-; If not $FE, return normally to draw count
-; Item is $FE - skip count but don't clear
-    bra _t_skip_to_rts
-
-_t_clear_count:
-; Write $FF (blank tiles) to count area: colon + 2 digits = 3 tiles
-    lda #0xFF
-; Blank tile
-    sta (0x29), y
-; Colon position
-    iny
-    lda.b 0xdb
-; Attribute byte
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-; First digit position
-    iny
-    lda.b 0xdb
-    sta (0x29), y
-    iny
-    lda #0xFF
-    sta (0x29), y
-; Second digit position
-    iny
-    lda.b 0xdb
-    sta (0x29), y
-
-_t_skip_to_rts:
-; Replace the entire DrawItemSlot return chain so we land on the RTS at $01:A222
-; instead of falling back to $A205 (NOPs + colon write).
-;
-; Stack on entry to Impl (pushed by trampoline + caller, top first):
-;   [JSL Impl return: PCL, PCH, PB ($01)]
-;   [JSR check_and_clear_count return: PCL, PCH ($A204)]
-;   [DrawItemSlot caller's return ...]
-;
-; Pop both the JSL Impl return (3 bytes) AND the JSR trampoline return (2 bytes),
-; then push a synthetic return to $01:A222 (RTS) for our RTL.
-    pla  ; PCL of JSL Impl return
-    pla  ; PCH of JSL Impl return
-    pla  ; PB  of JSL Impl return
-    pla  ; PCL of JSR check_and_clear_count return
-    pla  ; PCH of JSR check_and_clear_count return
-; Stack top now = DrawItemSlot's own caller return.
-; Push return for RTL: $01:A221 (so RTL lands PC=$01:A222 = RTS).
-    lda #0x01
-    pha
-    lda #0xA2
-    pha
-    lda #0x21
-    pha
-
-_t_normal_return:
-    rtl
-
-treasure_circular_slot_calc:
-
-
-"""
-Calculate tilemap Y offset using circular buffer position.
-Called from patched code at $A1BA via CircularSlotCalc_ext.
-
-Input: $5D = game's slot counter (0, 2, 4, 6... incremented by 2 per row)
-Output: Y = tilemap offset for circular buffer slot
-Preserves: 16-bit A mode on exit
-"""
-
-
-    sep #0x20
-; 8-bit A
-; Check if circular buffer mode is active (HDMA enabled)
-    lda.l 0x7E0000 + treasure_hdma_enable
-    beq _t_circ_slot_original
-; Not active, use original calculation
-; Circular buffer Y calculation
-; NOTE: Game increments $5D by 2 for each row (0, 2, 4, 6, 8, 10, 12, 14, 16, 18)
-; We must divide by 2 first to get the visual slot (0-9)
-    lda.b 0x5d
-; Load slot counter (0, 2, 4...)
-    lsr
-; Divide by 2 to get visual slot (0-9)
-    clc
-    adc.l 0x7E0000 + treasure_rolling_buffer_pos
-; Add buffer_pos
-
-_t_circ_slot_mod:
-    cmp #TREASURE_BUFFER_SLOTS
-; >= 11?
-    bcc _t_circ_slot_done
-    sec
-    sbc #TREASURE_BUFFER_SLOTS
-; Subtract 11 to wrap
-    bra _t_circ_slot_mod
-
-_t_circ_slot_done:
-; A = circular slot (0-10)
-    rep #0x20
-; 16-bit A
-    and.w #0x00FF
-; Clear high byte (force 16-bit immediate)
-    xba
-; A = slot * 256
-    lsr
-; A = slot * 128
-    clc
-    adc.w #0x0004
-; + 4 (margin only, $29 already includes border)
-    tay
-; Y = tilemap offset
-    rts
-
-_t_circ_slot_original:
-; Original game calculation: Y = ($5D / 2) * 128 + 4
-    lda.b 0x5d
-    lsr
-; /2
-    rep #0x20
-; 16-bit A
-    and.w #0x00FF
-; Clear high byte (force 16-bit immediate)
-    xba
-; *256
-    lsr
-; *128
-    clc
-    adc.w #0x0004
-; +4
-    tay
-    rts
-
-treasure_circular_slot_calc_ext:
-"""Trampoline to call CircularSlotCalc from bank $01 patch at $A1BA"""
-    jsr.w treasure_circular_slot_calc
-    rtl
+.include "../bank20.i"
+
+.alloc treasure_rolling_block in bank20_reloc {
+    init_treasure_inventory_hdma:
+    """
+    Sets up HDMA channel 5 for per-scanline BG1 vertical scroll control
+    Called when entering the item menu
+
+    HDMA mode: $02 = write 2 bytes to same register, DIRECT mode (like FF6)
+    Register: $210E (BG1VOFS)
+    Table format: count_byte, lo_byte, hi_byte per entry, $00 to end
+    """
+
+
+        php
+        sep #0x20  ; 8-bit A
+
+    ; Build the HDMA data table in WRAM
+    ; Use UpdateMenuScrollHDMA directly - InitMenuHDMATable is redundant
+    ; since UpdateMenuScrollHDMA is called immediately after anyway
+        jsr.w update_treasure_scroll_hdma
+
+    ; Configure HDMA channel 6 for DIRECT mode
+    ; Must use long addressing - DB may be $7E but registers are at $00:43xx
+        lda #0x02  ; Mode: DIRECT, write 2 bytes to same PPU reg
+        sta.l TREASURE_HDMA6_CTRL  ; $004360
+
+        lda #0x12  ; BG3VOFS register ($2112) — treasure inventory is on BG3
+        sta.l TREASURE_HDMA6_DEST  ; $004361
+
+    ; Source = HDMA table in WRAM at $7E9800
+        rep #0x20  ; 16-bit A
+        lda.w #TREASURE_HDMA_TABLE_ADDR  ; $9800
+        sta.l TREASURE_HDMA6_SRC_LO  ; $004362-$004363
+        sep #0x20  ; 8-bit A
+        lda #TREASURE_HDMA_BANK  ; $7E
+        sta.l TREASURE_HDMA6_SRC_BANK  ; $004364
+
+    ; HDMA channel 5 is now enabled via shadow variable (treasure_hdma_enable)
+    ; The NMI hook at $8083 reads the shadow and writes to HDMAEN
+
+        plp
+        rts
+
+    disable_treasure_inventory_hdma:
+    """
+    Disables HDMA channel 5 when leaving item menu
+    The shadow variable is cleared by menu_exit_hook
+    """
+        php
+        sep #0x20  ; 8-bit A
+        ; Shadow variable cleared by caller, NMI will write 0 to HDMAEN
+        plp
+        rts
+
+    init_treasure_hdma_table:
+    """
+    Builds direct mode HDMA table in WRAM at $7E9800
+    Format: count, lo, hi per entry, $00 to end
+    Initial state: all rows at BASE scroll (no circular buffer offset)
+    """
+        rep #0x30  ; 16-bit A, X, Y
+        php
+        pha  ; Save A
+        phx  ; Save X
+        phy  ; Save Y
+
+    ; Save DP bytes we'll use as scratch
+        lda.b 0x40
+        pha  ; Save $40-$41
+
+    ; X = table write offset
+        ldx.w #0x0000
+
+    ; Entry 0: Border area - 48 scanlines at BASE scroll
+        sep #0x20  ; 8-bit A for count byte
+        lda #48  ; 48 scanlines
+        sta.l TREASURE_HDMA_TABLE, x
+        inx
+        rep #0x20  ; 16-bit A for value
+        lda.w treasure_rolling_base_scroll
+        sta.l TREASURE_HDMA_TABLE, x
+        inx
+        inx
+
+    ; Entries 1-10: Item rows - 16 scanlines each at BASE scroll (initial)
+    ; For init, all rows use BASE (no circular buffer offset yet)
+        lda.w #0x0000
+        sta.b 0x40  ; Row counter
+
+    _t_init_item_rows:
+        sep #0x20  ; 8-bit A for count
+        lda #16  ; 16 scanlines per item row
+        sta.l TREASURE_HDMA_TABLE, x
+        inx
+        rep #0x20  ; 16-bit A for value
+        lda.w treasure_rolling_base_scroll
+        sta.l TREASURE_HDMA_TABLE, x
+        inx
+        inx
+
+        inc.b 0x40
+        lda.b 0x40
+
+        cmp.w #1  ; 10 rows
+        bcc _t_init_item_rows
+
+    ; Entry 11: Below items - 16 scanlines at BASE + 16
+    ; Lock to show the bottom window border area
+        sep #0x20
+        lda #16  ; 16 scanlines
+        sta.l TREASURE_HDMA_TABLE, x
+        inx
+        rep #0x20
+        lda.w treasure_rolling_base_scroll
+        clc
+        adc.w #16  ; Lock at base + 16
+        sta.l TREASURE_HDMA_TABLE, x
+        inx
+        inx
+
+    ; End marker
+        sep #0x20
+        lda #0x00
+        sta.l TREASURE_HDMA_TABLE, x
+
+    ; Restore DP bytes
+        rep #0x20
+        pla
+        sta.b 0x40  ; Restore $40-$41
+
+    ; Restore registers
+        ply
+        plx
+        pla
+        plp
+        rts
+
+    update_treasure_scroll_hdma:
+    """Build the treasure-menu HDMA scroll table via the shared engine."""
+        ; Build treasure HDMA scroll table. Inlined from the former
+        ; `engine_update_scroll_hdma` macro since its per-row writes use
+        ; compile-time state_base + hdma_shadow_addr ; routing through the
+        ; bank-20 engine would need menu_id dispatch at every state read,
+        ; which costs more than the modest 80-line duplication across the
+        ; four menus.
+    {
+        php
+        rep #0x30
+        pha
+        phx
+        phy
+        lda.b 0x40
+        pha
+        lda.b 0x42
+        pha
+        ldx.w #0x0000
+        jsr.w _treasure_hdma_header
+        stz.b 0x42
+
+    _row_loop:
+        lda.w treasure_rolling + RollingBufferState.buffer_pos
+        and.w #0x00FF
+        clc
+        adc.b 0x42
+
+    _mod_loop:
+        cmp.w #TREASURE_BUFFER_SLOTS
+        bcc _mod_done
+        sec
+        sbc.w #TREASURE_BUFFER_SLOTS
+        bra _mod_loop
+
+    _mod_done:
+        asl
+        asl
+        asl
+        asl
+        sta.b 0x40
+        lda.b 0x42
+        and.w #0x00FF
+        asl
+        asl
+        asl
+        asl
+        eor.w #0xFFFF
+        inc
+        clc
+        adc.b 0x40
+        clc
+        adc.w treasure_rolling + RollingBufferState.base_scroll
+        sta.b 0x40
+        sep #0x20
+        lda #16
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.b 0x40
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        inx
+        rep #0x20
+        inc.b 0x42
+        lda.b 0x42
+        cmp.w #TREASURE_VISIBLE_ITEMS
+        bcs _row_loop_done
+        jmp.w _row_loop
+
+    _row_loop_done:
+        jsr.w _treasure_hdma_footer
+        sep #0x20
+        lda #0x00
+        sta.l TREASURE_HDMA_SHADOW, x
+        jsr.w _treasure_hdma_signal
+        rep #0x20
+        pla
+        sta.b 0x42
+        pla
+        sta.b 0x40
+        ply
+        plx
+        pla
+        plp
+        rts
+    }
+
+    _treasure_hdma_header:
+    """Treasure profile header: drops band (120 lines at BASE-16) + inventory border (8 lines at BASE)."""
+
+
+        sep #0x20
+        lda #120
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w treasure_rolling_base_scroll
+        sec
+        sbc.w #16
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        inx
+        sep #0x20
+        lda #8
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w treasure_rolling_base_scroll
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _treasure_hdma_footer:
+    """Treasure profile footer: 16 scanlines at BASE+16 (hides prefetch slot)."""
+
+
+        sep #0x20
+        lda #16
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        rep #0x20
+        lda.w treasure_rolling_base_scroll
+        clc
+        adc.w #16
+        sta.l TREASURE_HDMA_SHADOW, x
+        inx
+        inx
+        rts
+
+    _treasure_hdma_signal:
+    """Treasure profile signal: set both treasure flag and shared $1BB6 mirror."""
+        sep #0x20
+        lda #0x01
+        sta.w treasure_hdma_copy_pending
+        sta.w treasure_hdma_copy_pending_shared
+        rts
+
+    ; Re-render all BUFFER_SLOTS (6) entries from $1440 using current
+    ; $1BB7 (scroll_pos) and the existing buffer_pos rotation. Called from
+    ; the redraw helper at $01:D933 after a swap completes — the swap
+    ; mutates $1440 in place and original's DrawInventoryList would redraw
+    ; the whole 48-item list, but we only need to refresh the 5 visible
+    ; slots + 1 prefetch. Crucially does NOT touch buffer_pos or any of
+    ; the state-machine bytes, so the cursor/scroll position the user
+    ; was on before the swap survives.
+
+    treasure_refresh_slots_impl:
+    """Treasure profile: re-render all 6 slots via the bank-20 engine (post-swap redraw at $01:D933)."""
+        php
+        sep #0x20
+        rep #0x10
+        lda.l 0x7E0000 + 0x1BB7  ; treasure scroll_pos (shared with vanilla cursor row)
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_refresh_slots
+        plp
+        rtl
+
+    init_treasure_rolling_buffer_impl:
+    """
+    Init treasure rolling buffer (5 visible + prefetch slot 6) via the
+    bank-20 rolling-inventory engine. State + hook far-ptrs live at
+    $7E:9C00  ; the engine writes the config block then JSLs into
+    `rolling_engine.rolling_engine_init` to zero the scratch, stamp
+    `base_scroll = $FFFF`, fire draw_window + ensure_hdma hooks, and
+    render `visible_rows` slots.
+    """
+        php
+        rep #0x30
+        sep #0x20
+        lda.b #TREASURE_BUFFER_SLOTS
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.visible_rows
+        lda.b #0x02
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.slot_height_tiles
+        lda.b #0x40
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.item_list_ptr
+        lda.b #0x14
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.item_list_ptr + 1
+        lda.b #0x7E
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.item_list_ptr + 2
+        lda.b #TREASURE_TOTAL_ITEMS
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.item_count
+        lda.b #0x06
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.hdma_channel
+        lda.b #0x80
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.vwf_cfg_ptr
+        lda.b #0x70
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.vwf_cfg_ptr + 1
+        lda.b #0x70
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.vwf_cfg_ptr + 2
+        lda.b #treasure_fn_render_slot_trampoline & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_render_slot
+        lda.b #( treasure_fn_render_slot_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_render_slot + 1
+        lda.b #( treasure_fn_render_slot_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_render_slot + 2
+        lda.b #treasure_fn_update_hdma_trampoline & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_update_hdma
+        lda.b #( treasure_fn_update_hdma_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_update_hdma + 1
+        lda.b #( treasure_fn_update_hdma_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_update_hdma + 2
+        lda.b #treasure_fn_draw_window_trampoline & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_draw_window
+        lda.b #( treasure_fn_draw_window_trampoline >> 8 ) & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_draw_window + 1
+        lda.b #( treasure_fn_draw_window_trampoline >> 16 ) & 0xFF
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.fn_draw_window + 2
+        lda.b #ROLLING_MENU_ID_TREASURE
+        sta.l 0x7E0000 + treasure_rolling + RollingBufferState.menu_id
+        plp
+        php
+        rep #0x10
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_init
+        plp
+        rtl
+
+    treasure_fn_render_slot_trampoline:
+    """Bank-20 RTL wrapper around `_treasure_render_item_to_slot`."""
+        php
+        jsr.w _treasure_render_item_to_slot
+        plp
+        rtl
+
+    treasure_fn_update_hdma_trampoline:
+    """Bank-20 RTL wrapper around `treasure_ensure_hdma_initialized`."""
+        php
+        jsr.w treasure_ensure_hdma_initialized
+        plp
+        rtl
+
+    treasure_fn_draw_window_trampoline:
+    """Bank-20 RTL wrapper around `_treasure_draw_inventory_window`."""
+        php
+        jsr.w _treasure_draw_inventory_window
+        plp
+        rtl
+
+    _treasure_draw_inventory_window:
+    """Treasure menu draws a 5-row inventory window so the bottom border lands on the rolling-buffer footer scanlines."""
+        rep #0x10
+        ldy.w #treasure_inventory_window
+        jsr.l draw_window_trampoline
+        sep #0x10
+        rts
+
+    ; _treasure_render_item_to_slot
+    ; Renders an item to a specific circular buffer slot in the tilemap.
+    ;
+    ; Input: treasure_rolling_edge_row = item index (0-47) for data lookup
+    ;        treasure_rolling_slot_index = slot index (0-11) for destination
+    ;
+    ; Strategy: Set up $5d = slot_index (for Y position calculation),
+    ;           $5a = pointer to item data, then call game's DrawItemSlot.
+
+    _treasure_render_item_to_slot:
+        php
+        phb
+
+    ; Set data bank to $7E for WRAM access
+        lda #0x7E
+        pha
+        plb
+
+    ; Set 16-bit A and X/Y for consistent register handling
+        rep #0x30  ; 16-bit A and X/Y
+
+    ; Save registers and key direct page variables
+        pha
+        phx  ; Save X (16-bit)
+        phy  ; Save Y (16-bit)
+        lda.b 0x5a
+        pha
+        lda.b 0x29  ; Save tilemap buffer pointer
+        pha
+        lda.b 0x45  ; Save $45-$46 (used by game routines)
+        pha
+        lda.b 0x33  ; Save $33-$34 (used for tile attributes)
+        pha
+        sep #0x20  ; 8-bit A
+        lda.b 0x5d
+        pha
+        lda.b 0xDB  ; Save tile attribute byte
+        pha
+
+    ; Set $29 = $B600 for BG1 tilemap buffer
+        rep #0x20  ; 16-bit A (X/Y still 16-bit)
+        lda.w #0xD600  ; BG3 screen buffer (treasure inventory lives on BG3)
+        sta.b 0x29
+        sep #0x20  ; 8-bit A
+
+    ; Calculate item data pointer: $1440 + (edge_row * Item.__size)
+        lda.w treasure_rolling_edge_row
+        asl  ; * Item.__size (2 bytes per Item)
+        clc
+        adc #0x40  ; Low byte of $1440
+        sta.b 0x5a
+        lda #0x14  ; High byte of $1440
+        adc #0x00  ; Add carry
+        sta.b 0x5b
+
+    ; Load Item.id and Item.qty from ($5A) via long-addressing into WRAM.
+        rep #0x20
+        lda.b 0x5a  ; Pointer value = $1440 + edge_row * Item.__size
+        tax
+        sep #0x20
+        lda.l 0x7E0000 + Item.id, x
+        pha  ; Save Item.id for CheckCanUseItem
+        lda.l 0x7E0000 + Item.qty, x
+        sta.b 0x5C  ; Store qty in $5C
+
+    ; Call CheckCanUseItem to set palette in $DB
+    ; Input: A = item ID
+    ; Output: $DB = $00 (usable) or $04 (not usable)
+        stz.b 0x34  ; Clear $34 (no priority/flip bits)
+        pla  ; Restore item ID
+        jsr.l check_can_use_item_trampoline  ; bank-$01 trampoline for original @ $A25D (sets $DB)
+
+    ; Set $5d = slot_index (for AND #$01 check, but we patched to AND #$00)
+        lda.w treasure_rolling_slot_index
+        sta.b 0x5d
+
+    ; Calculate Y = slot_index * 128 + $44
+    ; Y is the tilemap offset for this slot
+    ; +64 = 1 BG row (top window border at staging row 0)
+    ; +4 for left margin (2 tiles)
+        rep #0x20  ; 16-bit A (X/Y already 16-bit)
+        lda.w treasure_rolling_slot_index
+        and.w #0x00FF  ; Clear high byte
+        xba  ; Swap bytes: A = slot * 256
+        lsr  ; A = slot * 128
+        clc
+        adc.w #0x0044  ; + 68 (64 border + 4 margin)
+        tay  ; Y = tilemap offset (16-bit transfer)
+        sep #0x20  ; 8-bit A
+
+    ; Check for trash can item ($FF) - needs special 2x2 tile graphic
+        lda (0x5a)  ; Load item ID
+        cmp #0xFF
+        bne _t_not_trash_item
+        jsr.w draw_trash_treasure  ; Draw trash icon
+        bra _t_skip_draw_item_slot
+
+    _t_not_trash_item:
+        ; Clear the 2x2 trash can area first (in case we scrolled from trash position)
+        jsr.w _clear_treasure_trash_area
+
+    ; Call DrawItemSlot inner at $A1ED
+    ; Expects: Y = tilemap offset, ($5A) = item pointer, ($29) = tilemap base
+        jsr.l draw_item_slot_inner_trampoline  ; bank-$01 trampoline for original @ $A1ED
+
+    _t_skip_draw_item_slot:
+
+    ; Restore direct page variables and registers (reverse order)
+        pla
+        sta.b 0xDB  ; Restore tile attribute byte
+        pla
+        sta.b 0x5d
+        rep #0x20  ; 16-bit A
+        pla
+        sta.b 0x33  ; Restore $33-$34 (tile attributes)
+        pla
+        sta.b 0x45  ; Restore $45-$46 (used by game routines)
+        pla
+        sta.b 0x29  ; Restore tilemap buffer pointer
+        pla
+        sta.b 0x5a
+        rep #0x10  ; 16-bit X/Y for pop (match push)
+        ply  ; Restore Y (16-bit)
+        plx  ; Restore X (16-bit)
+        pla  ; Restore A (16-bit - still in 16-bit A from above)
+
+        plb
+        plp
+        rts
+
+    treasure_ensure_hdma_initialized:
+    """
+    Lazy initialization: captures $93 and sets up HDMA on first scroll.
+    Called from scroll prepare functions.
+    Checks if base_scroll == 0xFFFF (sentinel) and if so, initializes.
+    """
+
+    ; Check if already initialized (base_scroll != 0xFFFF)
+        rep #0x20  ; 16-bit A
+        lda.w treasure_rolling_base_scroll
+        cmp.w #0xFFFF
+        bne _t_hdma_already_init
+
+    ; Capture original BG3VOFS shadow ($9F). Original treasure menu uses
+    ; $9F = -120 to position items at screen scanline 120; we mirror that.
+        lda.l 0x7E019F
+        sta.w treasure_rolling_base_scroll
+        ; Clear the held-DOWN debounce counter ; engine_init_rolling_buffer
+        ; zeros the 12-byte engine struct but cooldown lives one byte past,
+        ; so explicitly nuke it here so the very first scroll trigger fires
+        ; immediately after popup-open.
+        sep #0x20
+        lda #0x00
+        sta.w treasure_scroll_cooldown
+        rep #0x20
+
+    ; Initialize HDMA channel configuration
+        sep #0x20  ; Back to 8-bit for InitMenuInventoryHDMA
+        jsr.w init_treasure_inventory_hdma
+
+    ; Enable HDMA via the SHARED field-menu shadow at $1BAE. The existing
+    ; field NMI hook (`field_menu_nmi_dma_transfer_check_impl`) reads this
+    ; byte and copies the HDMA shadow→active table each frame.
+    ; Treasure-only `treasure_hdma_enable` ($1BD6) is kept as a tracking
+    ; flag but isn't read by the NMI path.
+    ; OR-in ch6 enable bit ($40) so drops's ch4 bit ($10) — set by
+    ; drops_ensure_hdma_initialized at menu open — survives. Plain
+    ; `sta` would clobber the drops enable; same lazy-init order issue
+    ; as treasure_force_hdma_setup which uses $F9 (= ch7|ch6|ch5|ch4|ch3|ch0).
+        lda.l 0x7E1BAE
+        ora #0x40  ; Channel 6 enable (treasure rolling buffer)
+        sta.l 0x7E1BAE
+        sta.w treasure_hdma_enable
+        ; Push the enable straight to HDMAEN ($420C) so the very first frame
+        ; after popup-open has ch6 active. Without this immediate write, the
+        ; field NMI hook only copies $1BAE -> $420C on the NEXT vblank, which
+        ; left the first rendered frame with HDMA off and BG3VOFS at whatever
+        ; the previous menu left in $93 / $9F. Treasure inventory items then
+        ; rendered at the WRONG vertical scanline and looked like garbled
+        ; stride for a frame before settling.
+        lda.l 0x00420C
+        ora #0x40
+        sta.l 0x00420C
+        rts
+
+    _t_hdma_already_init:
+        sep #0x20  ; Restore 8-bit mode
+        rts
+
+    ; STATE MACHINE ROUTINES (FF6-style non-blocking scroll)
+
+    treasure_scroll_state_check:
+    """
+    Called at main loop entry ($019FF2) to handle scroll animation frames.
+    If scrolling is active, processes one frame and skips input handling.
+
+    Returns: Carry clear = process input normally
+     Carry set = skip input (still scrolling)
+    """
+
+
+        php
+        sep #0x20  ; 8-bit A
+
+    ; Tick cooldown counter so the next held-DOWN scroll is debounced.
+        lda.w treasure_scroll_cooldown
+        beq _t_scroll_cd_done
+        dec.w treasure_scroll_cooldown
+
+    _t_scroll_cd_done:
+        ; Check if we're scrolling
+        lda.w treasure_scroll_state
+        beq _t_scroll_state_idle
+
+    ; We're scrolling - process one animation frame
+        jsr.l treasure_update_scroll_frame_impl
+
+    ; Check if scroll finished
+        lda.w treasure_scroll_remaining
+        bne _t_scroll_still_active
+
+    ; Scroll finished - clean up and return to idle
+        jsr.l treasure_finish_scroll_impl
+
+    _t_scroll_state_idle:
+        plp
+        clc  ; Carry clear = process input
+        rts
+
+    _t_scroll_still_active:
+        plp
+        sec  ; Carry set = skip input
+        rts
+
+    treasure_start_scroll_down_impl:
+    """Treasure profile: kick scroll-down state machine via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1BB7  ; treasure scroll_pos (shared with vanilla cursor)
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_down
+        plp
+        rtl
+
+    treasure_start_scroll_up_impl:
+    """Treasure profile: kick scroll-up state machine via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1BB7
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_start_scroll_up
+        plp
+        rtl
+
+    treasure_update_scroll_frame_impl:
+    """Treasure profile: per-frame scroll animation tick."""
+        php
+        rep #0x10
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_update_scroll_frame
+        plp
+        rtl
+
+    treasure_finish_scroll_impl:
+    """Treasure profile: end-of-animation cleanup via the bank-20 engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1BB7
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_finish_scroll
+        plp
+        rtl
+
+    ; Inventory Rolling Buffer Patches - Relocated to Bank $20
+    ; These routines are called via JSL from bank $01 hooks.
+
+    .if INVENTORY_ROLLING_BUFFER {
+    treasure_menu_entry_hook_impl:
+    """Treasure profile: menu-entry implementation (HDMA capture + flush)."""
+        stz.w 0x1B1F
+        lda #0x00
+        sta.l 0x7E1BAE
+    ; treasure_hdma_enable
+        stz.w treasure_scroll_state
+        stz.w treasure_scroll_remaining
+        stz.w treasure_scroll_direction
+        stz.w treasure_transfer_pending
+        stz.w treasure_hdma_copy_pending
+    ; Clear HDMA copy flag
+        stz.w treasure_scroll_anim_offset
+    ; Clear low byte
+        stz.w treasure_scroll_anim_offset + 1
+    ; Clear high byte
+    ; Initialize cursor column to 0 for single-column mode
+    ; This ensures $1b22 is always 0 even if it had a value from previous menu
+        stz.w 0x1B22
+    ; cursor_x = 0
+    ; InitMenuRollingBuffer_Impl is called later via patched JSR at $9F7B
+        rtl
+
+    treasure_menu_exit_hook_impl:
+    """Treasure menu teardown: zero HDMA shadow, full 12-byte state, ch6 regs, original $1BC6 flag."""
+        php
+        sep #0x20
+        lda #0x00
+        sta.l 0x7E1BAE
+    ; treasure_hdma_enable shadow off
+        sta.l 0x7E1BC6
+    ; restore original "in treasure menu" flag (was original `stz $1BC6` at $01:D7E6 before the hook patch)
+        sta.l 0x004360
+        sta.l 0x004361
+        sta.l 0x004362
+        sta.l 0x004363
+        sta.l 0x004364
+    ; HDMA6 ctrl/dest/src cleared
+        rep #0x20
+        lda.w #0x0000
+        sta.w treasure_rolling
+        sta.w treasure_rolling + 2
+        sta.w treasure_rolling + 4
+        sta.w treasure_rolling + 6
+        sta.w treasure_rolling + 8
+        sta.w treasure_rolling + 10
+        plp
+        jsr.l reset_sprites_trampoline
+        rtl
+    ; treasure_swap_redraw_hook_impl_body
+    ; Called after item swap to redraw visible items correctly.
+    ; Must render to the correct circular buffer slots based on current buffer_pos.
+    ; Does NOT reset buffer_pos - we stay at the current scroll position.
+
+    treasure_swap_redraw_hook_impl_body:
+    """Treasure profile: post-swap re-render of all 6 slots via the engine."""
+        php
+        rep #0x10
+        lda.l 0x7E1BB7
+        ldx.w #treasure_rolling
+        jsr.l rolling_engine.rolling_engine_swap_redraw
+        plp
+        rtl
+
+
+    ; _clear_treasure_slot
+    ; Clears a single inventory slot in the tilemap buffer.
+    ; Input: treasure_rolling_slot_index = slot to clear (0-10)
+    ; Used when item index is out of bounds (>= 48)
+
+    _clear_treasure_slot:
+        php
+        phb
+    ; Set data bank to $7E for WRAM access
+        lda #0x7E
+        pha
+        plb
+    ; Save registers - use 16-bit mode for consistent push/pop
+        rep #0x30
+    ; 16-bit A and X/Y
+        pha
+        phx
+        phy
+        lda.b 0x29
+        pha
+    ; Set $29 = $B600 for BG1 tilemap buffer
+        lda.w #0xD600  ; BG3 screen buffer (treasure inventory lives on BG3)
+        sta.b 0x29
+    ; Calculate Y = slot_index * 128 + 70
+        lda.w treasure_rolling_slot_index
+        and.w #0x00FF
+        xba
+    ; A = slot * 256
+        lsr
+    ; A = slot * 128
+        clc
+        adc.w #0x0044
+    ; + 68 (64 border + 4 margin)
+        tay
+    ; Y = tilemap offset (16-bit)
+        sep #0x20
+    ; 8-bit A for tile writes (X/Y stay 16-bit)
+    ; Clear the item name area (12 tiles = 24 bytes)
+    ; Use $FF as blank tile
+    ; Note: X is 16-bit but we only use low byte; loop works correctly
+        ldx.w #12
+    ; 12 tiles for item name + quantity (force 16-bit immediate)
+
+    _t_clear_slot_loop:
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+        iny
+        lda #0x04
+    ; Palette 4 (matches normal items)
+        sta (0x29), y
+        iny
+        dex
+        bne _t_clear_slot_loop
+    ; Restore registers - must match push mode
+        rep #0x20
+    ; 16-bit A for pop (X/Y already 16-bit)
+        pla
+        sta.b 0x29
+        ply
+        plx
+        pla
+        plb
+        plp
+        rts
+    ; _clear_treasure_trash_area
+    ; Clears the 2x2 trash can area with blank tiles ($FF).
+    ; Called before drawing normal items to remove any leftover trash icon.
+    ; Input: Y = tilemap offset
+    ;        ($29) = tilemap base
+    ;
+
+    _clear_treasure_trash_area:
+        phy
+    ; Save Y
+    ; First row: 2 tiles
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+    ; Second row: Y + 64 from start
+        ply
+    ; Restore original Y
+        phy
+    ; Save again
+        rep #0x20
+        tya
+        clc
+        adc.w #64
+        tay
+        sep #0x20
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+        iny
+        lda #0x00
+        sta (0x29), y
+        ply
+    ; Restore Y
+        rts
+
+    draw_trash_treasure:
+
+
+    """
+    Draws the trash can 2x2 tile graphic for single-column inventory.
+    Input: Y = tilemap offset (from slot calculation)
+    ($29) = tilemap base ($B600)
+    treasure_rolling_slot_index = current slot
+    Tiles: $04 (top-left), $05 (top-right), $06 (bottom-left), $07 (bottom-right)
+
+    Tilemap format: [tile_number, attributes] pairs
+    Each row is 64 bytes (32 tiles × 2 bytes)
+    """
+
+
+    ; Y points to start of item slot area
+    ; Draw 2x2 trash can icon, then clear remaining 10 tiles per row
+    ; Save starting Y for second row calculation
+        rep #0x20
+    ; 16-bit for push
+        phy
+    ; Save starting Y
+        sep #0x20
+    ; 8-bit A
+    ; First row: tiles $04, $05
+        lda #0x04
+    ; Tile $04 (top-left of trash can)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute: palette 0, no flip
+        sta (0x29), y
+        iny
+        lda #0x05
+    ; Tile $05 (top-right)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+    ; Clear remaining 13 tiles on first row (10 name + colon + 2 digits)
+        ldx.w #13
+
+    _t_clear_row1:
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+        dex
+        bne _t_clear_row1
+    ; Restore starting Y and add 64 for second row
+        rep #0x20
+    ; 16-bit A
+        pla
+    ; Get starting Y
+        clc
+        adc.w #64
+    ; +64 bytes = next tilemap row
+        tay
+        sep #0x20
+    ; 8-bit A
+    ; Second row: tiles $06, $07
+        lda #0x06
+    ; Tile $06 (bottom-left)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+        lda #0x07
+    ; Tile $07 (bottom-right)
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+    ; Clear remaining 13 tiles on second row (10 name + colon + 2 digits)
+        ldx.w #13
+
+    _t_clear_row2:
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+        iny
+        lda #0x00
+    ; Attribute
+        sta (0x29), y
+        iny
+        dex
+        bne _t_clear_row2
+        rts
+    ; NmiDmaTransferCheck_Impl moved to bank $20 (battle/inventory_rolling.s)
+    ; as field_menu_nmi_dma_transfer_check_impl to save space in bank $01
+    ; treasure_check_and_clear_count_impl
+    ; Called from DrawItemSlot to check item ID and handle count display.
+    ; - If item ID is 0: writes $FF tiles to clear count, skips to RTS
+    ; - If item ID is $FE: skips to RTS (no clearing needed)
+    ; - Otherwise: returns normally to draw count
+    ;
+    ; Input: $5a = pointer to item data, Y = tilemap offset, $29 = tilemap ptr
+    ; Modifies: A, return address on stack if skipping
+
+    treasure_check_and_clear_count_impl:
+    """Treasure profile: drop the count column for empty/used slots (impl)."""
+        lda (0x5a)
+    ; Load item ID
+        beq _t_clear_count
+    ; If 0, clear and skip
+        cmp #0xFE
+    ; Check for special item $FE
+        bne _t_normal_return
+    ; If not $FE, return normally to draw count
+    ; Item is $FE - skip count but don't clear
+        bra _t_skip_to_rts
+
+    _t_clear_count:
+    ; Write $FF (blank tiles) to count area: colon + 2 digits = 3 tiles
+        lda #0xFF
+    ; Blank tile
+        sta (0x29), y
+    ; Colon position
+        iny
+        lda.b 0xdb
+    ; Attribute byte
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+    ; First digit position
+        iny
+        lda.b 0xdb
+        sta (0x29), y
+        iny
+        lda #0xFF
+        sta (0x29), y
+    ; Second digit position
+        iny
+        lda.b 0xdb
+        sta (0x29), y
+
+    _t_skip_to_rts:
+    ; Replace the entire DrawItemSlot return chain so we land on the RTS at $01:A222
+    ; instead of falling back to $A205 (NOPs + colon write).
+    ;
+    ; Stack on entry to Impl (pushed by trampoline + caller, top first):
+    ;   [JSL Impl return: PCL, PCH, PB ($01)]
+    ;   [JSR check_and_clear_count return: PCL, PCH ($A204)]
+    ;   [DrawItemSlot caller's return ...]
+    ;
+    ; Pop both the JSL Impl return (3 bytes) AND the JSR trampoline return (2 bytes),
+    ; then push a synthetic return to $01:A222 (RTS) for our RTL.
+        pla  ; PCL of JSL Impl return
+        pla  ; PCH of JSL Impl return
+        pla  ; PB  of JSL Impl return
+        pla  ; PCL of JSR check_and_clear_count return
+        pla  ; PCH of JSR check_and_clear_count return
+    ; Stack top now = DrawItemSlot's own caller return.
+    ; Push return for RTL: $01:A221 (so RTL lands PC=$01:A222 = RTS).
+        lda #0x01
+        pha
+        lda #0xA2
+        pha
+        lda #0x21
+        pha
+
+    _t_normal_return:
+        rtl
+
+    treasure_circular_slot_calc:
+
+
+    """
+    Calculate tilemap Y offset using circular buffer position.
+    Called from patched code at $A1BA via CircularSlotCalc_ext.
+
+    Input: $5D = game's slot counter (0, 2, 4, 6... incremented by 2 per row)
+    Output: Y = tilemap offset for circular buffer slot
+    Preserves: 16-bit A mode on exit
+    """
+
+
+        sep #0x20
+    ; 8-bit A
+    ; Check if circular buffer mode is active (HDMA enabled)
+        lda.l 0x7E0000 + treasure_hdma_enable
+        beq _t_circ_slot_original
+    ; Not active, use original calculation
+    ; Circular buffer Y calculation
+    ; NOTE: Game increments $5D by 2 for each row (0, 2, 4, 6, 8, 10, 12, 14, 16, 18)
+    ; We must divide by 2 first to get the visual slot (0-9)
+        lda.b 0x5d
+    ; Load slot counter (0, 2, 4...)
+        lsr
+    ; Divide by 2 to get visual slot (0-9)
+        clc
+        adc.l 0x7E0000 + treasure_rolling_buffer_pos
+    ; Add buffer_pos
+
+    _t_circ_slot_mod:
+        cmp #TREASURE_BUFFER_SLOTS
+    ; >= 11?
+        bcc _t_circ_slot_done
+        sec
+        sbc #TREASURE_BUFFER_SLOTS
+    ; Subtract 11 to wrap
+        bra _t_circ_slot_mod
+
+    _t_circ_slot_done:
+    ; A = circular slot (0-10)
+        rep #0x20
+    ; 16-bit A
+        and.w #0x00FF
+    ; Clear high byte (force 16-bit immediate)
+        xba
+    ; A = slot * 256
+        lsr
+    ; A = slot * 128
+        clc
+        adc.w #0x0004
+    ; + 4 (margin only, $29 already includes border)
+        tay
+    ; Y = tilemap offset
+        rts
+
+    _t_circ_slot_original:
+    ; Original game calculation: Y = ($5D / 2) * 128 + 4
+        lda.b 0x5d
+        lsr
+    ; /2
+        rep #0x20
+    ; 16-bit A
+        and.w #0x00FF
+    ; Clear high byte (force 16-bit immediate)
+        xba
+    ; *256
+        lsr
+    ; *128
+        clc
+        adc.w #0x0004
+    ; +4
+        tay
+        rts
+
+    treasure_circular_slot_calc_ext:
+    """Trampoline to call CircularSlotCalc from bank $01 patch at $A1BA"""
+        jsr.w treasure_circular_slot_calc
+        rtl
+    }
 }
diff --git a/src/intro.s b/src/intro.s
index ab05895..3b69dd5 100644
--- a/src/intro.s
+++ b/src/intro.s
@@ -8,123 +8,127 @@ and chains into the title screen.
 ; ----------------------------------------------------------------
 
 .include "libmz.i"  ; macros (dma_transfer_to_vram_call, etc.)
+.include "bank20.i"
 .import "libmz"
 .import "assets"
 .extern clear_ram
 
-start_splash_screen:
-"""Boot-time splash-screen entry point."""
-    ; initialise SNES
-    jsr.w initialize_snes
-
-; set register modes
-    rep #0x10  ; make X & Y 16-bits
-    sep #0x20  ; make A 8-bits
-
-; initialise graphics hardware
-    lda #0x03  ; graphics mode 3
-    sta 0x2105
-    lda #0x01  ; enable plane 0
-    sta 0x212c
-    lda #0x00  ; set plane 0 memory to 0x0000, 32x32 chars
-    sta 0x2107
-    lda #0x01  ; set plane 0 character set to 0x1000
-    sta 0x210b
-
-; copy intro map data
-    dma_transfer_to_vram_call(assets_intro_map, 0x0000, assets_intro_map__size, 0x1801)
-
-; copy color palettes
-    dma_transfer_to_palette_call(assets_intro_col, assets_intro_col__size)
-
-; copy intro tile set
-    dma_transfer_to_vram_call(assets_intro_set, 0x1000, assets_intro_set__size, 0x1801)
-
-    jsr.w _splash_screen_fade_in
-
-    lda #0x80
-    jsr.w _gamepad_interruptable_loop
-
-    jsr.w _splash_screen_fade_out
-
-    jsr.l clear_ram
-    ; runs the original jsl routines
-    jsr.l 0x15C8DF
-    jsr.l 0x15C9AA
-
-    rtl
-
-; TODO: rewrite as HDMA table would make it look less hacky.
-
-_splash_screen_fade_out:
-{
-    stz 0x00
-loop:
-    inc 0x00
-    lda #0x0F
-    sbc 0x00
-    sta 0x2100
-    lda 0x00
-    cmp #0x0F
-    beq exit
-    lda 0x00
-    asl
-    asl
-    asl
-    asl
-    inc
-    sta 0x2106
-
-    jsr.w wait_for_vblank
-    jsr.w wait_for_vblank
-    jsr.w wait_for_vblank
-    bra loop
-exit:
-    rts
-}
-
-
-_splash_screen_fade_in:
-{
-    stz 0x00
-loop:
-    inc 0x00
-    lda 0x00
-    sta 0x2100
-    asl
-    asl
-    asl
-    asl
-    sta 0x01
-    lda #0xF0
-    sec
-    sbc 0x01
-    inc
-    sta 0x2106
-
-    jsr.w wait_for_vblank
-    jsr.w wait_for_vblank
-
-    lda 0x00
-    cmp #0x0F
-    beq exit
-    bra loop
-exit:
-    rts
-}
-
-
-_gamepad_interruptable_loop:
-    ; 8bit A: Number of iterations
-{
-    jsr.w enable_gamepad
-    jsr.w wait_for_vblank
 
-    ldx 0x4218  ; lecture depuis joystick
-    bne exit  ; si on appuye sur quelque chose on sort du delay
-    dec
-    bne _gamepad_interruptable_loop
-exit:
-    jsr.w disable_gamepad
-    rts
+.alloc intro_block in bank20_reloc {
+    start_splash_screen:
+    """Boot-time splash-screen entry point."""
+        ; initialise SNES
+        jsr.w initialize_snes
+
+    ; set register modes
+        rep #0x10  ; make X & Y 16-bits
+        sep #0x20  ; make A 8-bits
+
+    ; initialise graphics hardware
+        lda #0x03  ; graphics mode 3
+        sta 0x2105
+        lda #0x01  ; enable plane 0
+        sta 0x212c
+        lda #0x00  ; set plane 0 memory to 0x0000, 32x32 chars
+        sta 0x2107
+        lda #0x01  ; set plane 0 character set to 0x1000
+        sta 0x210b
+
+    ; copy intro map data
+        dma_transfer_to_vram_call(assets_intro_map, 0x0000, assets_intro_map__size, 0x1801)
+
+    ; copy color palettes
+        dma_transfer_to_palette_call(assets_intro_col, assets_intro_col__size)
+
+    ; copy intro tile set
+        dma_transfer_to_vram_call(assets_intro_set, 0x1000, assets_intro_set__size, 0x1801)
+
+        jsr.w _splash_screen_fade_in
+
+        lda #0x80
+        jsr.w _gamepad_interruptable_loop
+
+        jsr.w _splash_screen_fade_out
+
+        jsr.l clear_ram
+        ; runs the original jsl routines
+        jsr.l 0x15C8DF
+        jsr.l 0x15C9AA
+
+        rtl
+
+    ; TODO: rewrite as HDMA table would make it look less hacky.
+
+    _splash_screen_fade_out:
+    {
+        stz 0x00
+    loop:
+        inc 0x00
+        lda #0x0F
+        sbc 0x00
+        sta 0x2100
+        lda 0x00
+        cmp #0x0F
+        beq exit
+        lda 0x00
+        asl
+        asl
+        asl
+        asl
+        inc
+        sta 0x2106
+
+        jsr.w wait_for_vblank
+        jsr.w wait_for_vblank
+        jsr.w wait_for_vblank
+        bra loop
+    exit:
+        rts
+    }
+
+
+    _splash_screen_fade_in:
+    {
+        stz 0x00
+    loop:
+        inc 0x00
+        lda 0x00
+        sta 0x2100
+        asl
+        asl
+        asl
+        asl
+        sta 0x01
+        lda #0xF0
+        sec
+        sbc 0x01
+        inc
+        sta 0x2106
+
+        jsr.w wait_for_vblank
+        jsr.w wait_for_vblank
+
+        lda 0x00
+        cmp #0x0F
+        beq exit
+        bra loop
+    exit:
+        rts
+    }
+
+
+    _gamepad_interruptable_loop:
+        ; 8bit A: Number of iterations
+    {
+        jsr.w enable_gamepad
+        jsr.w wait_for_vblank
+
+        ldx 0x4218  ; lecture depuis joystick
+        bne exit  ; si on appuye sur quelque chose on sort du delay
+        dec
+        bne _gamepad_interruptable_loop
+    exit:
+        jsr.w disable_gamepad
+        rts
+    }
 }
diff --git a/src/items.i b/src/items.i
index 4138348..6c00e25 100644
--- a/src/items.i
+++ b/src/items.i
@@ -1,19 +1,23 @@
-
-
 """
-Standard FF4 inventory item layout: 2-byte (id, qty) pairs.
+Inventory item layout + VWF region map shared across field, treasure,
+drops, and key-item menus. Defines the `Item` struct (2-byte id + qty
+pair) plus the BG3 CHR tile-id windows each menu surface owns.
+"""
 
-Used by every menu surface that shows player items in single-column
+.include "src/vwf_state.i"
 
-rolling-buffer form:
-  - Field menu Items submenu       at $7E:1440 (48 slots)
-  - Treasure menu inventory list   at $7E:1440 (same 48-slot array)
-  - Treasure menu drops list       at $7E:FF28 (8 drop slots)
-  - Key-item picker filter buffer  at $7E:0712 (filtered subset of $1440)
+; Standard FF4 inventory item layout: 2-byte (id, qty) pairs.
+;
+; Used by every menu surface that shows player items in single-column
 
-Battle inventory at $321A is a separate 4-byte layout (flags + id +
-qty + spell) and gets its own struct in a follow-up plan.
-"""
+; rolling-buffer form:
+;   - Field menu Items submenu       at $7E:1440 (48 slots)
+;   - Treasure menu inventory list   at $7E:1440 (same 48-slot array)
+;   - Treasure menu drops list       at $7E:FF28 (8 drop slots)
+;   - Key-item picker filter buffer  at $7E:0712 (filtered subset of $1440)
+;
+; Battle inventory at $321A is a separate 4-byte layout (flags + id +
+; qty + spell) and gets its own struct in a follow-up plan.
 
 
 ; Per-item record size in the French-translated assets_items_dat table:
@@ -23,6 +27,88 @@ qty + spell) and gets its own struct in a follow-up plan.
 ITEM_NAME_RECORD_SIZE := 0x0C
 ITEM_NAME_TEXT_SIZE := 0x0B
 
+; Per-item record size in the unleashed (16-char) name table at
+; `assets_items_unleashed_dat`: 1-byte symbol prefix + 16-byte name =
+; 17 bytes total. Battle / field / drops / treasure inventory all
+; render from this table once the BATTLE_ITEMS_VWF + field-menu
+; switches are on. Keep stride math + loop counters consistent via
+; these two symbols.
+ITEM_UNLEASHED_RECORD_SIZE := 0x11
+ITEM_UNLEASHED_TEXT_SIZE := 0x10
+
+; --- Field-menu BG3 CHR base ---
+; Menu PPU runs Mode 0 with BG34NBA = $22 (`ff4decomp/menu/menu.asm:3878+`),
+; mapping BG3 CHR to VRAM word $2000 / byte $4000. 256 static-font
+; tile_ids live at $00..$FF ; the VWF region starts past that at
+; tile_id $100 (9-bit tile_id territory) so glyph blits never
+; trample the menu chrome / font CHR.
+FIELD_BG3_CHR_VRAM_BYTE := 0x4000
+FIELD_BG3_CHR_VRAM_WORD := FIELD_BG3_CHR_VRAM_BYTE >> 1
+
+; VWF window starts at tile_id $100 in CHR = $4000 + $100 * 16 = $5000.
+FIELD_VWF_VRAM_DEST_BYTE := FIELD_BG3_CHR_VRAM_BYTE + 0x1000
+FIELD_VWF_VRAM_DEST_WORD := FIELD_VWF_VRAM_DEST_BYTE >> 1
+
+; --- Field-menu VWF regions ---
+; Mirrors the battle-side `region_size * N` partition in
+; `src/battle/message.s`. Each region owns a contiguous tile-id
+; window inside the BG3 CHR slot at VRAM $4000 (menu BG34NBA = $22).
+; Adding a new VWF surface (item description, treasure list,
+; whatever lives on field) means declaring a new region with its
+; tile-id base + budget + VRAM dest and the engine routes via
+; VwfConfig without touching the existing regions.
+;
+;   Region 0  $000..$0FF  vanilla menu font CHR (untouched)
+;   Region 1  $100..$169  field / treasure inventory item-name buffer
+;                         (FIELD_ITEM_VWF_TILE_BASE + K * slot,
+;                          11 slots * K=10 tile_ids)
+;   Region 1B $16E..$1A9  drops item-name buffer (treasure popup only),
+;                         6 buffer slots offset by DROPS_VWF_TILE_SLOT_OFFSET
+;                         (=11) from the treasure base so the two
+;                         panels coexist without clobbering each other.
+;
+; Field menu vs treasure-popup vs drops are UI-exclusive at the panel
+; level, but inside the treasure popup the treasure inventory AND the
+; monster-drops band render every frame. They both route through the
+; patched `DrawItemName` JSL hook in `items_menu_vwf.draw_field_item_name`
+; which computes tile_id_base from DP $5D. Drops's render hook offsets
+; $5D by +11 before calling the vanilla trampoline chain so its CHR
+; lands at $16E.. instead of $100.. ; items_menu_vwf bumps chr_byte_count
+; from $700 to $B00 so the NMI DMA flush covers both regions across
+; all 6 buffer slots per panel.
+;
+; Pushing the VWF region into 9-bit tile_id territory ($100+) keeps
+; it disjoint from the static font window ; the combined treasure +
+; drops footprint fits at $100..$1A9 (margin to $1FF before the next
+; BG3 CHR boundary).
+FIELD_ITEM_VWF_TILE_BASE := 0x100
+FIELD_ITEM_VWF_TILE_BUDGET := 0x0A
+DROPS_VWF_TILE_SLOT_OFFSET := 0x0B
+
+; --- Drops VWF flush descriptor (secondary NMI flush slot) ---
+; Hardcoded since drops only ever lives at the +11-slot offset in
+; the field BG3 CHR window. NMI runs both primary (treasure) and
+; secondary (drops) flushes per frame so each panel's CHR reaches
+; VRAM intact regardless of which descriptor was written last.
+;   src offset =  $16E   * 16 = $16E0  bytes into VWF_CHR_BUFFER
+;   vram dest  = ($5000 + $6E0) / 2 = $2B70 (= drops's CHR base)
+;   size       =  6 buffer slots * K=10 * 16 = $3C0 bytes
+DROPS_VWF_CHR_SRC_OFFSET := 0x16E0
+DROPS_VWF_VRAM_DEST_WORD := 0x2B70
+DROPS_VWF_BYTE_COUNT := 0x03C0
+
+; Primary flush size: covers the WORST-CASE primary consumer which
+; is the field-items menu (11 buffer slots * K=10 * 16 bytes = $6E0,
+; round up to $700). Treasure uses only 6 slots ($3C0) so the
+; primary DMA overshoots its region by a few tiles, but the extra
+; bytes flush into the start of drops's CHR window ($16E..$170)
+; whose secondary DMA overwrites them with the correct drops CHR
+; immediately afterwards. Was briefly $400 (treasure-only sized)
+; after the two-descriptor split, which cut field-items' last 5
+; buffer slots out of the flush window and tripped a BRK during
+; field scroll once tilemap entries pointed at unflushed tile_ids.
+FIELD_VWF_PRIMARY_BYTE_COUNT := 0x0700
+
 
 .struct Item {
     byte id
@@ -52,4 +138,20 @@ ITEM_NAME_TEXT_SIZE := 0x0B
     byte transfer_pending
     word scroll_anim_offset
     byte hdma_copy_pending
+    byte visible_rows
+    byte slot_height_tiles
+    long item_list_ptr
+    byte item_count
+    byte hdma_channel
+    long vwf_cfg_ptr
+    byte dirty_mask
+    long fn_render_slot
+    long fn_update_hdma
+    long fn_draw_window
+    byte menu_id
 }
+
+ROLLING_MENU_ID_FIELD := 0
+ROLLING_MENU_ID_TREASURE := 1
+ROLLING_MENU_ID_DROPS := 2
+ROLLING_MENU_ID_KEY_ITEM := 3
diff --git a/src/kerning.s b/src/kerning.s
index 64f2646..8e0afee 100644
--- a/src/kerning.s
+++ b/src/kerning.s
@@ -3,109 +3,113 @@ Dialog-VWF kerning helpers: long-form (RTL) wrapper around the binary-search bac
 kerning offsets.
 """
 .include "src/vwf.i"
+.include "bank20.i"
 
-dialog_get_kerning_adjustment_binary_search_ext:
-"""
-Long-form (RTL) wrapper around `dialog_get_kerning_adjustment_binary_search` for cross-bank callers and Python
-tests.
-"""
-    jsr.w dialog_get_kerning_adjustment_binary_search
-    rtl
 
+.alloc kerning_block in bank20_reloc {
+    dialog_get_kerning_adjustment_binary_search_ext:
+    """
+    Long-form (RTL) wrapper around `dialog_get_kerning_adjustment_binary_search` for cross-bank callers and Python
+    tests.
+    """
+        jsr.w dialog_get_kerning_adjustment_binary_search
+        rtl
 
-dialog_get_kerning_adjustment_binary_search:
-"""
-Binary-search the dialog font's kerning table for the pair in DP $25 (CURRENT_C)  ; returns adjustment in A or
-A=0+sec on miss. Trashes A/Y/flags, preserves X.
-"""
-    phx
-    jsr.w _dialog_get_kerning_adjustment_binary_search
-    plx
-    rts
 
-_dialog_get_kerning_adjustment_binary_search:
-{
-; Space ($FF) never appears in any font's kerning pair table; bail
-; before reading the count when either side of the pair is a space.
-; Caller is in 16-bit M.
-    lda.b CURRENT_C
-    and.w #0x00ff
-    cmp.w #0x00ff
-    beq _return_zero
-    lda.b CURRENT_C
-    and.w #0xff00
-    cmp.w #0xff00
-    beq _return_zero
+    dialog_get_kerning_adjustment_binary_search:
+    """
+    Binary-search the dialog font's kerning table for the pair in DP $25 (CURRENT_C)  ; returns adjustment in A or
+    A=0+sec on miss. Trashes A/Y/flags, preserves X.
+    """
+        phx
+        jsr.w _dialog_get_kerning_adjustment_binary_search
+        plx
+        rts
+
+    _dialog_get_kerning_adjustment_binary_search:
+    {
+    ; Space ($FF) never appears in any font's kerning pair table; bail
+    ; before reading the count when either side of the pair is a space.
+    ; Caller is in 16-bit M.
+        lda.b CURRENT_C
+        and.w #0x00ff
+        cmp.w #0x00ff
+        beq _return_zero
+        lda.b CURRENT_C
+        and.w #0xff00
+        cmp.w #0xff00
+        beq _return_zero
 
-    ldy.w #0x1100
-    lda.b [font_addr], y  ; NumKerningPairs (16-bit)
-    beq _return_zero
+        ldy.w #0x1100
+        lda.b [font_addr], y  ; NumKerningPairs (16-bit)
+        beq _return_zero
 
-    sec
-    sbc.w #0x0001  ; high = count - 1
-    pha  ; push high  (bottom)
-    lda.w #0x0000
-    pha  ; push low   (middle)
-    pea.w 0x0000  ; reserve mid slot (top)
+        sec
+        sbc.w #0x0001  ; high = count - 1
+        pha  ; push high  (bottom)
+        lda.w #0x0000
+        pha  ; push low   (middle)
+        pea.w 0x0000  ; reserve mid slot (top)
 
-_loop:
-    lda 0x0005, s  ; high
-    cmp 0x0003, s  ; compare against low
-    bcc _not_found  ; high < low -> no match
+    _loop:
+        lda 0x0005, s  ; high
+        cmp 0x0003, s  ; compare against low
+        bcc _not_found  ; high < low -> no match
 
-    lda 0x0005, s
-    sec
-    sbc 0x0003, s
-    lsr
-    clc
-    adc 0x0003, s
-    sta 0x0001, s  ; mid
+        lda 0x0005, s
+        sec
+        sbc 0x0003, s
+        lsr
+        clc
+        adc 0x0003, s
+        sta 0x0001, s  ; mid
 
-    lda 0x0001, s
-    asl
-    clc
-    adc 0x0001, s  ; mid * 3
-    clc
-    adc.w #0x1102
-    tay
+        lda 0x0001, s
+        asl
+        clc
+        adc 0x0001, s  ; mid * 3
+        clc
+        adc.w #0x1102
+        tay
 
-    lda.b [font_addr], y  ; load kerning key
-    cmp.b CURRENT_C
-    beq _found
-    bcc _search_upper  ; table key < target -> raise low
+        lda.b [font_addr], y  ; load kerning key
+        cmp.b CURRENT_C
+        beq _found
+        bcc _search_upper  ; table key < target -> raise low
 
-_search_lower:
-    lda 0x0001, s  ; mid - 1 becomes new high
-    sec
-    sbc.w #0x0001
-    bcc _not_found  ; mid was 0, underflow → not found
-    sta 0x0005, s
-    bra _loop
+    _search_lower:
+        lda 0x0001, s  ; mid - 1 becomes new high
+        sec
+        sbc.w #0x0001
+        bcc _not_found  ; mid was 0, underflow → not found
+        sta 0x0005, s
+        bra _loop
 
-_search_upper:
-    lda 0x0001, s  ; mid + 1 becomes new low
-    clc
-    adc.w #0x0001
-    sta 0x0003, s
-    bra _loop
+    _search_upper:
+        lda 0x0001, s  ; mid + 1 becomes new low
+        clc
+        adc.w #0x0001
+        sta 0x0003, s
+        bra _loop
 
-_not_found:
-    pla  ; discard mid
-    pla  ; discard low
-    pla  ; discard high
-_return_zero:
-    lda.w #0x0000
-    sec
-    rts
+    _not_found:
+        pla  ; discard mid
+        pla  ; discard low
+        pla  ; discard high
+    _return_zero:
+        lda.w #0x0000
+        sec
+        rts
 
-_found:
-    iny
-    iny
-    lda.b [font_addr], y
-    and.w #0x00FF
-    ply  ; drop mid  (PLY/X preserve A)
-    ply  ; drop low
-    ply  ; drop high
-    clc
-    rts
+    _found:
+        iny
+        iny
+        lda.b [font_addr], y
+        and.w #0x00FF
+        ply  ; drop mid  (PLY/X preserve A)
+        ply  ; drop low
+        ply  ; drop high
+        clc
+        rts
+    }
 }
diff --git a/src/lib/rolling_buffer.s b/src/lib/rolling_buffer.s
index e113c56..3ae3d84 100644
--- a/src/lib/rolling_buffer.s
+++ b/src/lib/rolling_buffer.s
@@ -28,542 +28,3 @@ Conventions:
 ;   settle is counted.
 INVENTORY_SCROLL_PIXELS_PER_FRAME := 8
 INVENTORY_SCROLL_TOTAL_PIXELS := 16
-
-
-.macro engine_update_scroll_hdma(
-    state_base,
-    hdma_shadow_addr,
-    buffer_slots,
-    visible,
-    header_hook,
-    footer_hook,
-    signal_hook,
-) {
-    """
-    Build the rolling-buffer HDMA scroll table in `hdma_shadow_addr`, based on
-    `state_base.buffer_pos` / `base_scroll`. Profile supplies header_hook
-    (above item rows), footer_hook (below item rows), signal_hook (NMI shadow
-    -> active copy pending).
-    """
-    php
-    rep #0x30
-    pha
-    phx
-    phy
-; Save DP scratch ($40-$43)
-    lda.b 0x40
-    pha
-    lda.b 0x42
-    pha
-    ldx.w #0x0000
-; Profile-specific top-of-table band(s).
-    jsr.w header_hook
-; Per-row entries: vram_slot = (buffer_pos + row) % buffer_slots
-    stz.b 0x42
-_row_loop:
-    lda.w state_base + RollingBufferState.buffer_pos
-    and.w #0x00FF
-    clc
-    adc.b 0x42
-_mod_loop:
-    cmp.w #buffer_slots
-    bcc _mod_done
-    sec
-    sbc.w #buffer_slots
-    bra _mod_loop
-_mod_done:
-; A = vram_slot (0..buffer_slots-1)
-    asl
-    asl
-    asl
-    asl
-; A = vram_slot * 16
-    sta.b 0x40
-; row * 16 = scanline_offset
-    lda.b 0x42
-    and.w #0x00FF
-    asl
-    asl
-    asl
-    asl
-; -scanline_offset
-    eor.w #0xFFFF
-    inc
-    clc
-    adc.b 0x40
-    clc
-    adc.w state_base + RollingBufferState.base_scroll
-    sta.b 0x40
-    sep #0x20
-    lda #16
-    sta.l hdma_shadow_addr, x
-    inx
-    rep #0x20
-    lda.b 0x40
-    sta.l hdma_shadow_addr, x
-    inx
-    inx
-    rep #0x20
-    inc.b 0x42
-    lda.b 0x42
-    cmp.w #visible
-    bcs _row_loop_done
-    jmp.w _row_loop
-_row_loop_done:
-; Profile-specific bottom-of-table band.
-    jsr.w footer_hook
-; End marker
-    sep #0x20
-    lda #0x00
-    sta.l hdma_shadow_addr, x
-; Profile-specific NMI signal (sets copy_pending shadow).
-    jsr.w signal_hook
-    rep #0x20
-    pla
-    sta.b 0x42
-    pla
-    sta.b 0x40
-    ply
-    plx
-    pla
-    plp
-    rts
-}
-
-
-.macro engine_init_rolling_buffer(state_base, render_count, draw_window_hook, ensure_hdma_hook, render_slot_hook) {
-    """
-    Initialise the rolling buffer on menu entry: zero the 12-byte state block,
-    mark base_scroll = $FFFF (sentinel), draw the inventory window border,
-    run the profile's `ensure_hdma_hook`, then render `render_count` slots via
-    `render_slot_hook` (each call sees `slot_index` already populated).
-    Field menu uses VISIBLE_ITEMS  ; treasure uses BUFFER_SLOTS.
-    """
-    php
-    pha
-    jsr.w draw_window_hook
-    lda.b 0x46
-    pha
-; Zero 12-byte rolling state, then mark base_scroll = $FFFF sentinel
-    rep #0x20
-    lda.w #0x0000
-    sta.w state_base
-    sta.w state_base + 2
-    sta.w state_base + 4
-    sta.w state_base + 6
-    sta.w state_base + 8
-    sta.w state_base + 10
-    lda.w #0xFFFF
-    sta.w state_base + RollingBufferState.base_scroll
-    sep #0x20
-    jsr.w ensure_hdma_hook
-    lda #0x00
-    sta.b 0x46
-_init_loop:
-    lda.b 0x46
-    sta.w state_base + RollingBufferState.edge_row
-    sta.w state_base + RollingBufferState.slot_index
-    jsr.w render_slot_hook
-    inc.b 0x46
-    lda.b 0x46
-    cmp #render_count
-    bne _init_loop
-    pla
-    sta.b 0x46
-    pla
-    plp
-    rtl
-}
-
-
-.macro engine_scroll_down_prepare(
-    state_base,
-    scroll_pos_addr,
-    scroll_limit,
-    visible,
-    buffer_slots,
-    ensure_hdma_hook,
-    render_slot_hook,
-    update_hdma_hook,
-) {
-    """
-    Pre-render the new bottom item before a scroll-down animation begins. The slot vacated at top (buffer_pos)
-    is reused for the new bottom  ; buffer_pos then advances modulo buffer_slots.
-    """
-    php
-    sep #0x20
-    jsr.w ensure_hdma_hook
-    lda.w scroll_pos_addr
-    cmp #scroll_limit
-    bcs _down_done
-    clc
-    adc #visible
-    sta.w state_base + RollingBufferState.edge_row
-    lda.w state_base + RollingBufferState.buffer_pos
-    sta.w state_base + RollingBufferState.slot_index
-    jsr.w render_slot_hook
-    inc.w state_base + RollingBufferState.buffer_pos
-    lda.w state_base + RollingBufferState.buffer_pos
-    cmp #buffer_slots
-    bcc _down_buf_ok
-    stz.w state_base + RollingBufferState.buffer_pos
-_down_buf_ok:
-    jsr.w update_hdma_hook
-_down_done:
-    plp
-    rts
-}
-
-
-.macro engine_scroll_up_prepare(
-    state_base,
-    scroll_pos_addr,
-    buffer_slots,
-    ensure_hdma_hook,
-    render_slot_hook,
-    update_hdma_hook,
-) {
-    """
-    Pre-render the new top item after the scroll-up state machine has decremented scroll_pos. buffer_pos walks
-    backwards (with wrap) so the slot leaving the bottom edge is reused for the new top.
-    """
-    php
-    sep #0x20
-    jsr.w ensure_hdma_hook
-    lda.w state_base + RollingBufferState.buffer_pos
-    beq _up_wrap
-    dec
-    bra _up_wrap_done
-_up_wrap:
-    lda #buffer_slots - 1
-_up_wrap_done:
-    sta.w state_base + RollingBufferState.buffer_pos
-    sta.w state_base + RollingBufferState.slot_index
-    lda.w scroll_pos_addr
-    sta.w state_base + RollingBufferState.edge_row
-    jsr.w render_slot_hook
-    jsr.w update_hdma_hook
-    plp
-    rts
-}
-
-
-.macro engine_start_scroll_down(
-    state_base,
-    scroll_pos_addr,
-    visible,
-    buffer_slots,
-    total_pixels,
-    pixels_per_frame,
-    ensure_hdma_hook,
-    render_slot_hook,
-    update_hdma_hook,
-) {
-    """
-    Kick off a non-blocking scroll-down animation: advance buffer_pos with wrap, pre-render the new bottom
-    item, configure the scroll state machine (-16 anim offset, +pixels_per_frame direction).
-    """
-    php
-    sep #0x20
-    jsr.w ensure_hdma_hook
-    inc.w state_base + RollingBufferState.buffer_pos
-    lda.w state_base + RollingBufferState.buffer_pos
-    cmp #buffer_slots
-    bcc _start_dn_buf_ok
-    stz.w state_base + RollingBufferState.buffer_pos
-    lda #0x00
-_start_dn_buf_ok:
-    clc
-    adc #visible - 1
-_start_dn_mod:
-    cmp #buffer_slots
-    bcc _start_dn_mod_done
-    sec
-    sbc #buffer_slots
-    bra _start_dn_mod
-_start_dn_mod_done:
-    sta.w state_base + RollingBufferState.slot_index
-    lda.w scroll_pos_addr
-    clc
-    adc #visible - 1
-    sta.w state_base + RollingBufferState.edge_row
-    jsr.w render_slot_hook
-    lda #0x01
-    sta.w state_base + RollingBufferState.scroll_state
-    lda #total_pixels
-    sta.w state_base + RollingBufferState.scroll_remaining
-    lda #pixels_per_frame
-    sta.w state_base + RollingBufferState.scroll_direction
-    rep #0x20
-    lda.w #0xFFF0
-    sta.w state_base + RollingBufferState.scroll_anim_offset
-    sep #0x20
-    lda #0x01
-    sta.w state_base + RollingBufferState.transfer_pending
-    jsr.w update_hdma_hook
-    plp
-    rtl
-}
-
-
-.macro engine_start_scroll_up(
-    state_base,
-    scroll_pos_addr,
-    buffer_slots,
-    total_pixels,
-    ensure_hdma_hook,
-    render_slot_hook,
-    update_hdma_hook,
-) {
-    """
-    Kick off a non-blocking scroll-up animation: walk buffer_pos backwards (wrap to buffer_slots-1),
-    pre-render the new top item, configure the state machine (+16 anim offset, -2 direction).
-    """
-    php
-    sep #0x20
-    jsr.w ensure_hdma_hook
-    lda.w state_base + RollingBufferState.buffer_pos
-    beq _start_up_wrap
-    dec
-    bra _start_up_wrap_done
-_start_up_wrap:
-    lda #buffer_slots - 1
-_start_up_wrap_done:
-    sta.w state_base + RollingBufferState.buffer_pos
-    sta.w state_base + RollingBufferState.slot_index
-    lda.w scroll_pos_addr
-    sta.w state_base + RollingBufferState.edge_row
-    jsr.w render_slot_hook
-    lda #0x01
-    sta.w state_base + RollingBufferState.scroll_state
-    lda #total_pixels
-    sta.w state_base + RollingBufferState.scroll_remaining
-    lda #0xFE
-    sta.w state_base + RollingBufferState.scroll_direction
-    rep #0x20
-    lda.w #0x0010
-    sta.w state_base + RollingBufferState.scroll_anim_offset
-    sep #0x20
-    lda #0x01
-    sta.w state_base + RollingBufferState.transfer_pending
-    jsr.w update_hdma_hook
-    plp
-    rtl
-}
-
-
-.macro engine_update_scroll_frame(state_base, pixels_per_frame, update_hdma_hook) {
-    """
-    One animation frame: advance scroll_anim_offset by pixels_per_frame towards 0 (sign-aware), nudge cursor
-    sprite at $0311 if 'second item' mode is active, decrement scroll_remaining, refresh HDMA, then push
-    sprites + BG2 to VRAM.
-    """
-    php
-    rep #0x20
-    lda.w state_base + RollingBufferState.scroll_anim_offset
-    sep #0x20
-    lda.w state_base + RollingBufferState.scroll_direction
-    bpl _frame_positive
-    rep #0x20
-    lda.w state_base + RollingBufferState.scroll_anim_offset
-    sec
-    sbc.w #pixels_per_frame
-    sta.w state_base + RollingBufferState.scroll_anim_offset
-    bra _frame_update_cursor
-_frame_positive:
-    rep #0x20
-    lda.w state_base + RollingBufferState.scroll_anim_offset
-    clc
-    adc.w #pixels_per_frame
-    sta.w state_base + RollingBufferState.scroll_anim_offset
-_frame_update_cursor:
-    sep #0x20
-    lda.w 0x1B19
-    beq _frame_no_cursor
-    lda.w state_base + RollingBufferState.scroll_direction
-    bpl _frame_cursor_down
-    inc.w 0x0311
-    inc.w 0x0311
-    bra _frame_no_cursor
-_frame_cursor_down:
-    dec.w 0x0311
-    dec.w 0x0311
-_frame_no_cursor:
-    lda.w state_base + RollingBufferState.scroll_remaining
-    sec
-    sbc #pixels_per_frame
-    sta.w state_base + RollingBufferState.scroll_remaining
-    jsr.w update_hdma_hook
-    jsr.l tfr_sprites_vblank_trampoline
-    jsr.l tfr_bg2_tiles_vblank_trampoline
-    plp
-    rtl
-}
-
-
-.macro engine_finish_scroll(
-    state_base,
-    scroll_pos_addr,
-    visible,
-    buffer_slots,
-    total_items,
-    render_slot_hook,
-    update_hdma_hook,
-) {
-    """
-    End-of-animation: pre-render the next-scroll-direction edge slot (skip past list ends), reset scroll_state
-    and anim_offset, refresh HDMA, run original cursor + post-scroll cleanup.
-    """
-    php
-    sep #0x20
-    lda.w state_base + RollingBufferState.scroll_direction
-    bmi _finish_was_up
-    lda.w state_base + RollingBufferState.buffer_pos
-    beq _finish_dn_wrap
-    dec
-    bra _finish_dn_slot_ok
-_finish_dn_wrap:
-    lda #buffer_slots - 1
-_finish_dn_slot_ok:
-    sta.w state_base + RollingBufferState.slot_index
-    lda.w scroll_pos_addr
-    clc
-    adc #visible
-    cmp #total_items
-    bcs _finish_skip
-    sta.w state_base + RollingBufferState.edge_row
-    jsr.w render_slot_hook
-    lda #0x01
-    sta.w state_base + RollingBufferState.transfer_pending
-    bra _finish_skip
-_finish_was_up:
-    lda.w state_base + RollingBufferState.buffer_pos
-    clc
-    adc #visible
-_finish_up_mod:
-    cmp #buffer_slots
-    bcc _finish_up_slot_ok
-    sec
-    sbc #buffer_slots
-    bra _finish_up_mod
-_finish_up_slot_ok:
-    sta.w state_base + RollingBufferState.slot_index
-    lda.w scroll_pos_addr
-    beq _finish_skip
-    dec
-    sta.w state_base + RollingBufferState.edge_row
-    jsr.w render_slot_hook
-    lda #0x01
-    sta.w state_base + RollingBufferState.transfer_pending
-_finish_skip:
-    stz.w state_base + RollingBufferState.scroll_state
-    rep #0x20
-    stz.w state_base + RollingBufferState.scroll_anim_offset
-    sep #0x20
-    jsr.w update_hdma_hook
-    jsr.l draw_item_cursors_trampoline
-    jsr.l update_ctrl_after_scroll_trampoline
-    plp
-    rtl
-}
-
-
-.macro engine_swap_redraw(
-    state_base,
-    scroll_pos_addr,
-    buffer_slots,
-    total_items,
-    ensure_hdma_hook,
-    render_slot_hook,
-    clear_slot_hook,
-    update_hdma_hook,
-) {
-    """
-    Re-render all `buffer_slots` slots from the current buffer_pos after an item swap. Out-of-range items
-    (scroll_pos+row >= total_items) are blanked via `clear_slot_hook` so the prefetch row doesn't show stale
-    tiles. Resets scroll state so a mid-animation swap doesn't trigger spurious finish_scroll re-renders.
-    """
-    php
-    sep #0x20
-    jsr.w ensure_hdma_hook
-    stz.w state_base + RollingBufferState.scroll_state
-    stz.w state_base + RollingBufferState.scroll_remaining
-    stz.w state_base + RollingBufferState.scroll_anim_offset
-    stz.w state_base + RollingBufferState.scroll_anim_offset + 1
-    lda.b 0x46
-    pha
-    lda #0x00
-    sta.b 0x46
-_swap_loop:
-    lda.w state_base + RollingBufferState.buffer_pos
-    clc
-    adc.b 0x46
-_swap_mod:
-    cmp #buffer_slots
-    bcc _swap_mod_done
-    sec
-    sbc #buffer_slots
-    bra _swap_mod
-_swap_mod_done:
-    sta.w state_base + RollingBufferState.slot_index
-    lda.w scroll_pos_addr
-    clc
-    adc.b 0x46
-    cmp #total_items
-    bcs _swap_clear
-    sta.w state_base + RollingBufferState.edge_row
-    jsr.w render_slot_hook
-    bra _swap_next
-_swap_clear:
-    jsr.w clear_slot_hook
-_swap_next:
-    inc.b 0x46
-    lda.b 0x46
-    cmp #buffer_slots
-    bne _swap_loop
-    pla
-    sta.b 0x46
-    lda #0x01
-    sta.w state_base + RollingBufferState.transfer_pending
-    jsr.w update_hdma_hook
-    plp
-    rtl
-}
-
-
-.macro engine_refresh_slots(state_base, scroll_pos_addr, buffer_slots, render_slot_hook) {
-    """
-    Re-render all `buffer_slots` from current buffer_pos without resetting scroll state. Used by treasure's
-    original redraw helper (post-swap rebuild path) where the engine itself didn't trigger the redraw.
-    """
-    php
-    rep #0x10
-    sep #0x20
-    lda #0x00
-    sta.b 0x46
-_refresh_loop:
-    lda.w scroll_pos_addr
-    clc
-    adc.b 0x46
-    sta.w state_base + RollingBufferState.edge_row
-    lda.w state_base + RollingBufferState.buffer_pos
-    clc
-    adc.b 0x46
-_refresh_mod:
-    cmp #buffer_slots
-    bcc _refresh_mod_done
-    sec
-    sbc #buffer_slots
-    bra _refresh_mod
-_refresh_mod_done:
-    sta.w state_base + RollingBufferState.slot_index
-    jsr.w render_slot_hook
-    inc.b 0x46
-    lda.b 0x46
-    cmp #buffer_slots
-    bne _refresh_loop
-    lda #0x01
-    sta.w state_base + RollingBufferState.transfer_pending
-    plp
-    rtl
-}
diff --git a/src/lib/rolling_inventory_engine.s b/src/lib/rolling_inventory_engine.s
new file mode 100644
index 0000000..190fb44
--- /dev/null
+++ b/src/lib/rolling_inventory_engine.s
@@ -0,0 +1,844 @@
+"""
+Unified rolling-inventory engine — bank-20 entry points.
+
+Phase 1 skeleton : declares the JSL-callable entry points specified in
+plans/rolling_inventory_engine.md. Each entry currently delegates to
+the existing macro-driven `engine_*` machinery in
+`src/lib/rolling_buffer.s`, so callers can opt-in incrementally.
+
+ABI : `X` (16-bit, bank $7E implied) holds the pointer to a
+`RollingBufferState` instance. Hooks live in bank-01 as far-callable
+trampolines and are resolved through the struct's per-instance far
+pointers (`vwf_cfg_ptr`, hook ptrs, ...).
+
+Phases 2..5 will port field-items / shop / key-items / drops + treasure
+through these entries one by one  ; phase 6 deletes the duplicated macro
+sites that remain in the per-menu source files.
+"""
+
+
+.include "../bank20.i"
+
+.alloc rolling_inventory_engine_block in bank20_reloc {
+    .scope rolling_engine {
+        """Bank-20 rolling-inventory engine entry points (phase 1 stubs)."""
+    rolling_engine_init:
+    """
+    Cold init for a rolling-inventory instance.
+
+    Zeroes the 12-byte engine scratch portion of `RollingBufferState`
+    (everything from top_row through transfer_pending plus the
+    scroll_anim_offset word + hdma_copy_pending byte) and stamps
+    `base_scroll = $FFFF` as the "not-yet-HDMA-initialised" sentinel
+    the per-menu ensure_hdma_hook checks. dirty_mask is cleared too so
+    the first vblank flush has nothing to fire.
+
+    Caller-managed config fields (visible_rows, slot_height_tiles,
+    item_list_ptr, item_count, hdma_channel, vwf_cfg_ptr) are NOT
+    touched  ; callers fill them in before invoking this entry and the
+    phase-2 render path will read them back.
+
+    In/out :
+      X = state ptr (16-bit, bank $7E implied)
+      Y = initial top_row (unused in phase 1  ; phase 2 will route this
+          into the scroll-position pre-render)
+      All registers clobbered.
+    """
+
+
+        {
+        php
+        sep #0x20
+        rep #0x10
+        lda.b #0x00
+    ; Zero engine scratch (top_row..transfer_pending + anim offset bytes +
+    ; hdma_copy_pending + dirty_mask).
+        sta.l 0x7E0000 + RollingBufferState.top_row, x
+        sta.l 0x7E0000 + RollingBufferState.buffer_pos, x
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        sta.l 0x7E0000 + RollingBufferState.hdma_enable, x
+    ; _pad byte (offset 7, struct-defined as `byte _pad`) ; a816 hides
+    ; leading-underscore field names from outer scopes, so write through
+    ; the literal offset rather than the symbolic name.
+        sta.l 0x7E0007, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_direction, x
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset + 1, x
+        sta.l 0x7E0000 + RollingBufferState.hdma_copy_pending, x
+        sta.l 0x7E0000 + RollingBufferState.dirty_mask, x
+    ; base_scroll = $FFFF sentinel ("HDMA not yet armed").
+        lda.b #0xFF
+        sta.l 0x7E0000 + RollingBufferState.base_scroll, x
+        sta.l 0x7E0000 + RollingBufferState.base_scroll + 1, x
+    ; Fire fn_draw_window hook when armed.
+        rep #0x10
+        ldy.w #RollingBufferState.fn_draw_window
+        jsr.w _engine_call_hook
+    ; Fire fn_update_hdma hook (= per-menu ensure_hdma_initialized in the
+    ; field-items wiring). Phase 3+ may split this into ensure-once vs
+    ; per-scroll-update hooks ; for now both share the slot.
+        ldy.w #RollingBufferState.fn_update_hdma
+        jsr.w _engine_call_hook
+    ; Loop `visible_rows` times, calling fn_render_slot for each row.
+    ; Each iteration plants edge_row + slot_index = loop counter before
+    ; firing the hook so the per-menu render reads the right values.
+        sep #0x20
+        lda.b #0x00
+
+    _engine_init_render_loop:
+        pha
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+        pla
+        inc
+        cmp.l 0x7E0000 + RollingBufferState.visible_rows, x
+        bcc _engine_init_render_loop
+
+        plp
+        rtl
+        }
+
+    rolling_engine_refresh_slots:
+    """
+    Re-render every slot in the rolling buffer from the current `buffer_pos`
+    without resetting scroll state. Mirrors the legacy `engine_refresh_slots`
+    macro : iterates `(visible_rows + 1)` times (visible window + prefetch
+    slot), each iteration setting edge_row = scroll_pos + counter and
+    slot_index = (buffer_pos + counter) % buffer_slots, then firing
+    fn_render_slot. Stamps `transfer_pending = 1` on exit so the per-menu
+    NMI hook DMAs the BG staging buffer next vblank.
+
+    In : X = state ptr (16-bit, bank $7E implied)
+         A = current scroll_pos value (8-bit)
+    Out: all `visible_rows + 1` slots re-rendered, transfer_pending set.
+    """
+
+
+        {
+        php
+        sep #0x20
+        rep #0x10
+        sta.l 0x001F88  ; stash scroll_pos
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+        inc  ; buffer_slots = visible_rows + 1
+        sta.l 0x001F89  ; stash buffer_slots
+        lda.b #0x00
+
+    _engine_refresh_loop:
+        pha
+        clc
+        adc.l 0x001F88
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        pla
+        pha
+        clc
+        adc.l 0x7E0000 + RollingBufferState.buffer_pos, x
+
+    _engine_refresh_mod:
+        cmp.l 0x001F89
+        bcc _engine_refresh_mod_done
+        sec
+        sbc.l 0x001F89
+        bra _engine_refresh_mod
+
+    _engine_refresh_mod_done:
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+        pla
+        inc
+        cmp.l 0x001F89
+        bcc _engine_refresh_loop
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        plp
+        rtl
+        }
+
+    rolling_engine_update_scroll_frame:
+    """
+    Advance the scroll animation one frame. Mirrors the legacy
+    `engine_update_scroll_frame` macro :
+      - scroll_anim_offset += (scroll_direction < 0 ? -8 : +8)
+      - if vanilla cursor-row marker $1B19 != 0, bump cursor sprite Y at
+        $0311 by 2 pixels in the matching direction (down for negative
+        direction, up for positive)
+      - scroll_remaining -= 8 (animation timer countdown)
+      - per-menu update_scroll_hdma trampoline rebuilds the HDMA shadow
+        table (dispatched via state.menu_id)
+      - tfr_sprites_vblank_trampoline + tfr_bg2_tiles_vblank_trampoline
+        push sprite OAM + BG2 tiles into VRAM so the animation frame
+        becomes visible.
+
+    Pixels-per-frame hardcoded at 8  ; every existing caller
+    (INVENTORY_SCROLL_PIXELS_PER_FRAME / TREASURE / DROPS / KEY_ITEM) uses
+    that value, so the constant moves into the engine.
+
+    In : X = state ptr (16-bit, bank $7E implied)
+    Out: scroll_anim_offset / scroll_remaining advanced, HDMA shadow +
+         sprite / BG2 VRAM uploads driven by the per-menu trampolines.
+    """
+
+
+        {
+        php
+        rep #0x10
+        sep #0x20
+        lda.l 0x7E0000 + RollingBufferState.scroll_direction, x
+        bpl _frame_positive
+        rep #0x20
+        lda.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        sec
+        sbc.w #8
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        bra _frame_update_cursor
+
+    _frame_positive:
+        rep #0x20
+        lda.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        clc
+        adc.w #8
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+
+    _frame_update_cursor:
+        sep #0x20
+        lda.l 0x7E1B19  ; vanilla cursor-row marker
+        beq _frame_no_cursor
+        lda.l 0x7E0000 + RollingBufferState.scroll_direction, x
+        bpl _frame_cursor_down
+        inc.w 0x0311
+        inc.w 0x0311
+        bra _frame_no_cursor
+
+    _frame_cursor_down:
+        dec.w 0x0311
+        dec.w 0x0311
+
+    _frame_no_cursor:
+        lda.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        sec
+        sbc.b #8
+        sta.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+    ; Dispatch per-menu update_scroll_hdma via menu_id branch. The four
+    ; profiles' update_*_scroll_hdma functions live in bank-20 alongside
+    ; this engine, so jsr.w (3-byte) reaches them cleanly.
+        lda.l 0x7E0000 + RollingBufferState.menu_id, x
+        beq _frame_hdma_field
+        cmp.b #ROLLING_MENU_ID_TREASURE
+        beq _frame_hdma_treasure
+        cmp.b #ROLLING_MENU_ID_DROPS
+        beq _frame_hdma_drops
+        jsr.w update_key_item_scroll_hdma
+        bra _frame_hdma_done
+
+    _frame_hdma_field:
+        jsr.w update_menu_scroll_hdma
+        bra _frame_hdma_done
+
+    _frame_hdma_treasure:
+        jsr.w update_treasure_scroll_hdma
+        bra _frame_hdma_done
+
+    _frame_hdma_drops:
+        jsr.w update_drops_scroll_hdma
+
+    _frame_hdma_done:
+        jsr.l tfr_sprites_vblank_trampoline
+        jsr.l tfr_bg2_tiles_vblank_trampoline
+        plp
+        rtl
+        }
+
+    rolling_engine_start_scroll_down:
+    """
+    Kick off a non-blocking scroll-down animation. Advances buffer_pos
+    with wrap, pre-renders the new bottom slot via fn_render_slot, configures
+    the scroll FSM (state = 1, remaining = 16 pixels, direction = +8 =
+    positive = down, anim_offset = -16), stamps transfer_pending, and
+    dispatches the per-menu update_scroll_hdma to seed the first frame's
+    HDMA shadow.
+
+    In : X = state ptr, A = current scroll_pos (8-bit)
+    Out: scroll state machine armed for the next 2 frames of animation.
+    """
+
+
+        {
+        php
+        rep #0x10
+        sep #0x20
+        sta.l 0x001F88  ; stash scroll_pos
+        ldy.w #RollingBufferState.fn_update_hdma
+        jsr.w _engine_call_hook
+    ; buffer_slots = visible_rows + 1
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+        inc
+        sta.l 0x001F89
+    ; inc buffer_pos with wrap
+        lda.l 0x7E0000 + RollingBufferState.buffer_pos, x
+        inc
+        cmp.l 0x001F89
+        bcc _start_dn_buf_ok
+        lda.b #0x00
+
+    _start_dn_buf_ok:
+        sta.l 0x7E0000 + RollingBufferState.buffer_pos, x
+    ; slot_index = (buffer_pos + visible_rows - 1) % buffer_slots
+        clc
+        adc.l 0x7E0000 + RollingBufferState.visible_rows, x
+        dec
+
+    _start_dn_mod:
+        cmp.l 0x001F89
+        bcc _start_dn_mod_done
+        sec
+        sbc.l 0x001F89
+        bra _start_dn_mod
+
+    _start_dn_mod_done:
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+    ; edge_row = scroll_pos + visible_rows - 1
+        lda.l 0x001F88
+        clc
+        adc.l 0x7E0000 + RollingBufferState.visible_rows, x
+        dec
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+    ; FSM : scroll_state = 1, remaining = 16, direction = +8 (positive = down)
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        lda.b #16
+        sta.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        lda.b #8
+        sta.l 0x7E0000 + RollingBufferState.scroll_direction, x
+        rep #0x20
+        lda.w #0xFFF0
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        sep #0x20
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        jsr.w _engine_dispatch_update_scroll_hdma
+        plp
+        rtl
+        }
+
+    rolling_engine_start_scroll_up:
+    """
+    Kick off a non-blocking scroll-up animation. Walks buffer_pos backwards
+    (wrap to visible_rows = buffer_slots - 1), pre-renders the new top
+    slot, configures the FSM (state = 1, remaining = 16, direction = -2 =
+    $FE = negative = up, anim_offset = +16), stamps transfer_pending,
+    dispatches update_scroll_hdma.
+
+    In : X = state ptr, A = current scroll_pos (8-bit, already decremented
+         by the caller)
+    Out: scroll state machine armed for next 2 frames of animation.
+    """
+
+
+        {
+        php
+        rep #0x10
+        sep #0x20
+        sta.l 0x001F88  ; stash scroll_pos
+        ldy.w #RollingBufferState.fn_update_hdma
+        jsr.w _engine_call_hook
+        lda.l 0x7E0000 + RollingBufferState.buffer_pos, x
+        beq _start_up_wrap
+        dec
+        bra _start_up_wrap_done
+
+    _start_up_wrap:
+    ; buffer_slots - 1 == visible_rows.
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+
+    _start_up_wrap_done:
+        sta.l 0x7E0000 + RollingBufferState.buffer_pos, x
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        lda.l 0x001F88  ; scroll_pos
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        lda.b #16
+        sta.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        lda.b #0xFE  ; -2 (negative = up direction)
+        sta.l 0x7E0000 + RollingBufferState.scroll_direction, x
+        rep #0x20
+        lda.w #0x0010
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        sep #0x20
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        jsr.w _engine_dispatch_update_scroll_hdma
+        plp
+        rtl
+        }
+
+    rolling_engine_finish_scroll:
+    """
+    End-of-animation : pre-render the next-direction edge slot (the one
+    that becomes the new prefetch after the current animation lands),
+    reset scroll_state + anim_offset, refresh the HDMA shadow, fire the
+    vanilla cursor + post-scroll cleanup trampolines.
+
+    For scroll-down (direction = +N): slot_index = buffer_pos - 1 (with
+    wrap to buffer_slots - 1)  ; edge_row = scroll_pos + visible_rows ;
+    skip render when scroll_pos + visible_rows >= item_count (past tail).
+
+    For scroll-up (direction < 0, $FE) : slot_index = (buffer_pos + visible)
+    mod buffer_slots  ; edge_row = scroll_pos - 1 ; skip render when
+    scroll_pos == 0 (past head).
+
+    In : X = state ptr, A = current scroll_pos (8-bit)
+    Out: scroll FSM cleared, prefetch slot rendered, HDMA refreshed.
+    """
+
+
+        {
+        php
+        rep #0x10
+        sep #0x20
+        sta.l 0x001F88  ; stash scroll_pos
+    ; buffer_slots = visible_rows + 1
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+        inc
+        sta.l 0x001F89
+        lda.l 0x7E0000 + RollingBufferState.scroll_direction, x
+        bmi _finish_was_up
+    ; --- scroll-down post-anim ---
+        lda.l 0x7E0000 + RollingBufferState.buffer_pos, x
+        beq _finish_dn_wrap
+        dec
+        bra _finish_dn_slot_ok
+
+    _finish_dn_wrap:
+    ; buffer_slots - 1 == visible_rows
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+
+    _finish_dn_slot_ok:
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        lda.l 0x001F88  ; scroll_pos
+        clc
+        adc.l 0x7E0000 + RollingBufferState.visible_rows, x
+        cmp.l 0x7E0000 + RollingBufferState.item_count, x
+        bcs _finish_skip
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        bra _finish_skip
+
+    _finish_was_up:
+    ; --- scroll-up post-anim ---
+        lda.l 0x7E0000 + RollingBufferState.buffer_pos, x
+        clc
+        adc.l 0x7E0000 + RollingBufferState.visible_rows, x
+
+    _finish_up_mod:
+        cmp.l 0x001F89
+        bcc _finish_up_slot_ok
+        sec
+        sbc.l 0x001F89
+        bra _finish_up_mod
+
+    _finish_up_slot_ok:
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        lda.l 0x001F88  ; scroll_pos
+        beq _finish_skip
+        dec
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+
+    _finish_skip:
+        lda.b #0x00
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        rep #0x20
+        lda.w #0x0000
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        sep #0x20
+        jsr.w _engine_dispatch_update_scroll_hdma
+        jsr.l draw_item_cursors_trampoline
+        jsr.l update_ctrl_after_scroll_trampoline
+        plp
+        rtl
+        }
+
+    rolling_engine_swap_redraw:
+    """
+    Re-render all `visible_rows + 1` slots from current buffer_pos after
+    an item swap. Slots whose item index >= item_count are cleared via
+    the built-in field-items clear path (other profiles' clear stubs were
+    RTS no-ops, so we skip the call entirely for them). Resets scroll
+    state so a mid-animation swap doesn't trigger spurious finish_scroll
+    re-renders.
+
+    In : X = state ptr, A = current scroll_pos (8-bit)
+    Out: all `visible_rows + 1` slots refreshed, transfer_pending = 1,
+         HDMA shadow rebuilt via per-menu update_scroll_hdma.
+    """
+
+
+        {
+        php
+        rep #0x10
+        sep #0x20
+        sta.l 0x001F88  ; stash scroll_pos
+        ldy.w #RollingBufferState.fn_update_hdma
+        jsr.w _engine_call_hook
+        lda.b #0x00
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_anim_offset + 1, x
+    ; buffer_slots = visible_rows + 1
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+        inc
+        sta.l 0x001F89  ; stash buffer_slots
+        lda.b #0x00
+
+    _swap_loop:
+        pha
+        clc
+        adc.l 0x7E0000 + RollingBufferState.buffer_pos, x
+
+    _swap_mod:
+        cmp.l 0x001F89
+        bcc _swap_mod_done
+        sec
+        sbc.l 0x001F89
+        bra _swap_mod
+
+    _swap_mod_done:
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        pla
+        pha
+        clc
+        adc.l 0x001F88
+        cmp.l 0x7E0000 + RollingBufferState.item_count, x
+        bcs _swap_clear
+        sta.l 0x7E0000 + RollingBufferState.edge_row, x
+        ldy.w #RollingBufferState.fn_render_slot
+        jsr.w _engine_call_hook
+        bra _swap_next
+
+    _swap_clear:
+    ; Only field-items has a real clear ; other menus' clear stubs were
+    ; RTS no-ops, so skip the call entirely unless menu_id == 0.
+        lda.l 0x7E0000 + RollingBufferState.menu_id, x
+        bne _swap_next
+        jsr.w _clear_inventory_slot
+
+    _swap_next:
+        pla
+        inc
+        cmp.l 0x001F89
+        bcc _swap_loop
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        jsr.w _engine_dispatch_update_scroll_hdma
+        plp
+        rtl
+        }
+
+    _engine_dispatch_update_scroll_hdma:
+    """
+    Branches to the per-menu update_*_scroll_hdma function based on
+    state.menu_id. All four functions live in bank-20 alongside the engine
+    so a 16-bit jsr.w is enough.
+    """
+
+
+        {
+        lda.l 0x7E0000 + RollingBufferState.menu_id, x
+        beq _dispatch_field
+        cmp.b #ROLLING_MENU_ID_TREASURE
+        beq _dispatch_treasure
+        cmp.b #ROLLING_MENU_ID_DROPS
+        beq _dispatch_drops
+        jsr.w update_key_item_scroll_hdma
+        rts
+
+    _dispatch_field:
+        jsr.w update_menu_scroll_hdma
+        rts
+
+    _dispatch_treasure:
+        jsr.w update_treasure_scroll_hdma
+        rts
+
+    _dispatch_drops:
+        jsr.w update_drops_scroll_hdma
+        rts
+        }
+
+    _engine_call_hook:
+    """
+    Internal helper. JSLs through a hook far-ptr stored at
+    `state[X + Y]`.
+
+    In  : X = state ptr (16-bit, bank $7E implied)
+          Y = struct offset of the 3-byte hook field
+          M = 8 (caller-managed)
+    Out : hook RTLs back to caller of _engine_call_hook. A clobbered.
+          Hook is skipped if all three bytes are zero  ; null-safe.
+    """
+
+
+        {
+        phx
+        phy
+    ; Stash hook-field offset at $00:1F83 (next to the $1F80-82 jmp-vec
+    ; cell). Direct-page `sty.b 0x90` looked tempting but FF4's menu code
+    ; sets DP=$0100 before calling our patch, so the store actually
+    ; landed at $0190 = the BG1HOFS shadow ; vanilla NMI then scrolled
+    ; BG1 by the hook offset (26 px for fn_render_slot) on every render.
+        rep #0x20
+        tya
+        sta.l 0x001F83
+        txa
+        clc
+        adc.l 0x001F83
+        tax  ; X now points at the hook field's first byte
+        sep #0x20
+        lda.l 0x7E0000, x
+        sta.l 0x001F80
+        lda.l 0x7E0001, x
+        sta.l 0x001F81
+        lda.l 0x7E0002, x
+        sta.l 0x001F82
+        ora.l 0x001F80
+        ora.l 0x001F81
+        beq _engine_call_hook_null
+        phk
+        rep #0x20
+        pea.w ( _engine_call_hook_return - 1 ) & 0xFFFF
+        sep #0x20
+        jmp [0x001F80]
+
+    _engine_call_hook_return:
+    _engine_call_hook_null:
+        ply
+        plx
+        rts
+        }
+
+    rolling_engine_tick:
+    """
+    Per-frame state-machine tick.
+
+    Advances the scroll animation by one frame when scroll_state != 0 :
+    decrements scroll_remaining and, when it hits zero, clears
+    scroll_state and stamps transfer_pending = 1 so the next vblank
+    flush picks up the final frame. Returns immediately when
+    scroll_state is idle.
+
+    Phase 1.3 scope. Phase 2 will additionally drive scroll_anim_offset,
+    read the pad-down bits in A.b to start fresh scrolls, and route the
+    new edge slot through fn_render_slot once that hook lands in the
+    struct.
+
+    In  : X = state ptr
+    Out : scroll_remaining decremented (if scrolling), scroll_state
+          cleared + transfer_pending set when animation completes.
+    """
+
+
+        {
+        php
+        sep #0x20
+        lda.l 0x7E0000 + RollingBufferState.scroll_state, x
+        beq _tick_idle
+        lda.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        sec
+        sbc.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.scroll_remaining, x
+        bne _tick_still_scrolling
+        lda.b #0x00
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        lda.b #0x01
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+
+    _tick_still_scrolling:
+    _tick_idle:
+        plp
+        rtl
+        }
+
+    rolling_engine_vblank_flush:
+    """
+    NMI-time flush.
+
+    Phase 1.4 scope : reads `state.dirty_mask`, returns immediately on
+    $00, otherwise clears the mask so the next vblank starts clean.
+    Phase 2 will wedge the real DMA-from-VWF-stage-to-VRAM dance in
+    between, driven by the `vwf_cfg_ptr` config and a small bank-01
+    trampoline  ; the dirty-mask gate + clear stays valid regardless of
+    how the DMA path lands.
+
+    Must run inside force-blank or the HDMA gap window so the DMA
+    doesn't tear visible scanlines (phase 2 concern).
+
+    In  : X = state ptr
+    Out : state.dirty_mask = 0 if it was non-zero, A clobbered.
+    """
+
+
+        {
+        php
+        sep #0x20
+        lda.l 0x7E0000 + RollingBufferState.dirty_mask, x
+        beq _flush_clean
+        lda.b #0x00
+        sta.l 0x7E0000 + RollingBufferState.dirty_mask, x
+
+    _flush_clean:
+        plp
+        rtl
+        }
+
+    rolling_engine_cursor_up:
+    """
+    Decrement slot_index, wrapping to (visible_rows - 1) at zero.
+
+    In  : X = state ptr (bank $7E implied)
+    Out : state.slot_index updated, A clobbered
+    """
+
+
+        {
+        php
+        sep #0x20
+        lda.l 0x7E0000 + RollingBufferState.slot_index, x
+        bne _cursor_up_dec
+        lda.l 0x7E0000 + RollingBufferState.visible_rows, x
+
+    _cursor_up_dec:
+        dec
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        plp
+        rtl
+        }
+
+    rolling_engine_cursor_down:
+    """
+    Increment slot_index, wrapping to 0 when it would equal visible_rows.
+
+    In  : X = state ptr
+    Out : state.slot_index updated, A clobbered
+    """
+
+
+        {
+        php
+        sep #0x20
+        lda.l 0x7E0000 + RollingBufferState.slot_index, x
+        inc
+        cmp.l 0x7E0000 + RollingBufferState.visible_rows, x
+        bcc _cursor_dn_store
+        lda.b #0x00
+
+    _cursor_dn_store:
+        sta.l 0x7E0000 + RollingBufferState.slot_index, x
+        plp
+        rtl
+        }
+
+    rolling_engine_invalidate_slot:
+    """
+    Mark one slot dirty for the next vblank flush.
+
+    In : X = state ptr, A.b = slot index 0..visible_rows-1
+    Out : state.dirty_mask gets bit (1 << A) set
+    """
+
+
+        {
+        php
+        sep #0x20
+        pha
+    ; X stays state ptr ; use Y for the shift-left tally so we can keep
+    ; X bound to the struct base for the dirty_mask write below.
+        phx
+        tay
+        lda.b #0x01
+
+    _inv_slot_shift:
+        cpy.w #0x0000
+        beq _inv_slot_done
+        asl
+        dey
+        bra _inv_slot_shift
+
+    _inv_slot_done:
+        plx
+        ora.l 0x7E0000 + RollingBufferState.dirty_mask, x
+        sta.l 0x7E0000 + RollingBufferState.dirty_mask, x
+        pla
+        plp
+        rtl
+        }
+
+    rolling_engine_invalidate_all:
+    """
+    Mark every visible slot dirty by setting dirty_mask = $FF.
+
+    In : X = state ptr
+    Out : state.dirty_mask = $FF, A clobbered
+    """
+
+
+        {
+        php
+        sep #0x20
+        lda.b #0xFF
+        sta.l 0x7E0000 + RollingBufferState.dirty_mask, x
+        plp
+        rtl
+        }
+
+    rolling_engine_shutdown:
+    """
+    Tear down state on menu close.
+
+    Zeroes `state.hdma_enable` + `state.dirty_mask` and resets
+    `state.base_scroll` to the $FFFF sentinel so the next cold-init
+    ensure_hdma_hook re-arms cleanly. Does not touch HDMAEN ($420C) :
+    the per-menu exit hook owns that bit (it knows which channel mask
+    to clear via `hdma_channel`).
+
+    In : X = state ptr
+    Out : engine state half-reset, A clobbered.
+    """
+
+
+        {
+        php
+        sep #0x20
+        rep #0x10
+        lda.b #0x00
+        sta.l 0x7E0000 + RollingBufferState.hdma_enable, x
+        sta.l 0x7E0000 + RollingBufferState.dirty_mask, x
+        sta.l 0x7E0000 + RollingBufferState.scroll_state, x
+        sta.l 0x7E0000 + RollingBufferState.transfer_pending, x
+        sta.l 0x7E0000 + RollingBufferState.hdma_copy_pending, x
+        lda.b #0xFF
+        sta.l 0x7E0000 + RollingBufferState.base_scroll, x
+        sta.l 0x7E0000 + RollingBufferState.base_scroll + 1, x
+        plp
+        rtl
+        }
+    }
+}
diff --git a/src/libmz.s b/src/libmz.s
index edcfd10..cfced55 100644
--- a/src/libmz.s
+++ b/src/libmz.s
@@ -1,237 +1,241 @@
 """Shared low-level helpers: vblank spinner + DMA transfer routines for VRAM / palette + small bank-trampolines."""
 
 .include "libmz.i"
+.include "bank20.i"
 
-wait_for_vblank:
-"""Spin until the next vblank edge: wait for $4212 to go low, then high."""
-{
-    pha
 
-_negative:
-    lda.l 0x004212
-    bmi _negative
+.alloc libmz_block in bank20_reloc {
+    wait_for_vblank:
+    """Spin until the next vblank edge: wait for $4212 to go low, then high."""
+    {
+        pha
 
-_positive:
-    lda.l 0x004212
-    bpl _positive
-    pla
-    rts
-}
+    _negative:
+        lda.l 0x004212
+        bmi _negative
 
-wait_for_vblank_long:
-"""RTL trampoline around `wait_for_vblank` for cross-bank callers."""
-    jsr.w wait_for_vblank
-    rtl
-
-dma_transfer_to_vram:
-"""
-DMA-transfer a block from ROM/RAM into VRAM via channel 7.
-Stack args (caller pushes in order): source_offset, source_bank,
-vram_pointer, count, dma_mode. All five args are pulled before RTS.
-"""
-{
-; on the stack:
-; return address
-; source offset
-; source bank
-; vram pointer
-; count
-; mode
-    arg_count = 5
-    stack_ptr = arg_count * 2 - 1
-    source_offset = stack_ptr
-    source_bank = stack_ptr - 2
-    vram_pointer = stack_ptr - 4
-    count = stack_ptr - 6
-    dma_mode = stack_ptr - 8
-    channel = 7
-    rep #0x20
-    sep #0x10
-    ldx #0x80
-    stx 0x2115
-    lda.b source_offset, s
-    sta.w 0x4302 + ( channel << 4 )
-    sep #0x10
-    lda.b source_bank, s
-    sta.w 0x4304 + ( channel << 4 )
-    rep #0x20
-    lda.b vram_pointer, s
-    sta.w 0x2116
-    lda.b count, s
-    sta.w 0x4305 + ( channel << 4 )
-    lda.b dma_mode, s
-    sta.w 0x4300 + ( channel << 4 )
-    ldx.b #1 << channel
-    stx 0x420B
-    nop
-    nop
-    pla
-    pla
-    pla
-    pla
-    pla
-    rts
-}
+    _positive:
+        lda.l 0x004212
+        bpl _positive
+        pla
+        rts
+    }
 
+    wait_for_vblank_long:
+    """RTL trampoline around `wait_for_vblank` for cross-bank callers."""
+        jsr.w wait_for_vblank
+        rtl
+
+    dma_transfer_to_vram:
+    """
+    DMA-transfer a block from ROM/RAM into VRAM via channel 7.
+    Stack args (caller pushes in order): source_offset, source_bank,
+    vram_pointer, count, dma_mode. All five args are pulled before RTS.
+    """
+    {
+    ; on the stack:
+    ; return address
+    ; source offset
+    ; source bank
+    ; vram pointer
+    ; count
+    ; mode
+        arg_count = 5
+        stack_ptr = arg_count * 2 - 1
+        source_offset = stack_ptr
+        source_bank = stack_ptr - 2
+        vram_pointer = stack_ptr - 4
+        count = stack_ptr - 6
+        dma_mode = stack_ptr - 8
+        channel = 7
+        rep #0x20
+        sep #0x10
+        ldx #0x80
+        stx 0x2115
+        lda.b source_offset, s
+        sta.w 0x4302 + ( channel << 4 )
+        sep #0x10
+        lda.b source_bank, s
+        sta.w 0x4304 + ( channel << 4 )
+        rep #0x20
+        lda.b vram_pointer, s
+        sta.w 0x2116
+        lda.b count, s
+        sta.w 0x4305 + ( channel << 4 )
+        lda.b dma_mode, s
+        sta.w 0x4300 + ( channel << 4 )
+        ldx.b #1 << channel
+        stx 0x420B
+        nop
+        nop
+        pla
+        pla
+        pla
+        pla
+        pla
+        rts
+    }
+
+
+    dma_transfer_to_palette:
+    """
+    DMA-transfer a block into CGRAM (palette) via channel 7.
+    Stack args (caller-pushed): source_offset, source_bank, count.
+    All three pulled before RTS.
+    """
+    {
+    ; on the stack:
+    ; return address
+    ; source offset
+    ; source bank
+    ; count
+        arg_count = 3
+        stack_ptr = arg_count * 2 - 1
+        source_offset = stack_ptr
+        source_bank = stack_ptr - 2
+        count = stack_ptr - 4
+        rep #0x20
+        sep #0x10
+        ldx #0x80
+        stx 0x2115
+        lda.b source_offset, s
+        sta 0x4372
+        lda.b source_bank, s
+        sta 0x4374
+        lda.b count, s
+        sta 0x4375
+        lda #0x2200
+        sta 0x4370
+        ldx #1 << 7
+        stx 0x420B
+        nop
+        nop
+        pla
+        pla
+        pla
+        rts
+    }
 
-dma_transfer_to_palette:
-"""
-DMA-transfer a block into CGRAM (palette) via channel 7.
-Stack args (caller-pushed): source_offset, source_bank, count.
-All three pulled before RTS.
-"""
-{
-; on the stack:
-; return address
-; source offset
-; source bank
-; count
-    arg_count = 3
-    stack_ptr = arg_count * 2 - 1
-    source_offset = stack_ptr
-    source_bank = stack_ptr - 2
-    count = stack_ptr - 4
-    rep #0x20
-    sep #0x10
-    ldx #0x80
-    stx 0x2115
-    lda.b source_offset, s
-    sta 0x4372
-    lda.b source_bank, s
-    sta 0x4374
-    lda.b count, s
-    sta 0x4375
-    lda #0x2200
-    sta 0x4370
-    ldx #1 << 7
-    stx 0x420B
-    nop
-    nop
-    pla
-    pla
-    pla
-    rts
-}
 
+    _enable_display:
+        pha
+        lda #0x00  ; enable screen, full brightness
+        sta 0x2100
+        pla
+        rts
 
-_enable_display:
-    pha
-    lda #0x00  ; enable screen, full brightness
-    sta 0x2100
-    pla
-    rts
-
-enable_gamepad:
-"""Enable auto-joypad-read + vblank NMI by writing 1 to $4200."""
-    pha
-    lda #0x01
-    sta 0x4200
-    pla
-    rts
-
-
-disable_gamepad:
-"""Disable auto-joypad-read and all NMI/IRQ sources ($4200 := 0)."""
-    stz 0x4200
-    rts
-
-
-initialize_snes:
-"""
-Reset PPU, DMA, and CPU registers to a known idle state at boot.
-8-bit M/X  ; clears OBJ/BG registers, neutralises mode-7 matrix to identity,
-disables NMI/HDMA/joypad, leaves screen forced-blank.
-"""
-    sep #0x30  ; make X, Y, A all 8-bits
-    lda #0x80  ; screen off, no brightness
-    sta 0x2100  ; brightness & screen enable register
-    lda #0x00
-    sta 0x2101  ; sprite register (size & address in VRAM)
-    sta 0x2102  ; sprite registers (address of sprite memory [OAM])
-    sta 0x2103  ; sprite registers (address of sprite memory [OAM])
-    sta 0x2105  ; graphic mode register
-    sta 0x2106  ; mosaic register
-    sta 0x2107  ; plane 0 map VRAM location
-    sta 0x2108  ; plane 1 map VRAM location
-    sta 0x2109  ; plane 2 map VRAM location
-    sta 0x210A  ; plane 3 map VRAM location
-    sta 0x210B  ; plane 0 & 1 Tile data location
-    sta 0x210C  ; plane 2 & 3 Tile data location
-    sta 0x210D  ; plane 0 scroll x (first 8 bits)
-    sta 0x210D  ; plane 0 scroll x (last 3 bits)
-    sta 0x210E  ; plane 0 scroll y (first 8 bits)
-    sta 0x210E  ; plane 0 scroll y (last 3 bits)
-    sta 0x210F  ; plane 1 scroll x (first 8 bits)
-    sta 0x210F  ; plane 1 scroll x (last 3 bits)
-    sta 0x2110  ; plane 1 scroll y (first 8 bits)
-    sta 0x2110  ; plane 1 scroll y (last 3 bits)
-    sta 0x2111  ; plane 2 scroll x (first 8 bits)
-    sta 0x2111  ; plane 2 scroll x (last 3 bits)
-    sta 0x2112  ; plane 2 scroll y (first 8 bits)
-    sta 0x2112  ; plane 2 scroll y (last 3 bits)
-    sta 0x2113  ; plane 3 scroll x (first 8 bits)
-    sta 0x2113  ; plane 3 scroll x (last 3 bits)
-    sta 0x2114  ; plane 3 scroll y (first 8 bits)
-    sta 0x2114  ; plane 3 scroll y (last 3 bits)
-    lda #0x80  ; increase VRAM address after writing to 0x2119
-    sta 0x2115  ; VRAM address increment register
-    lda #0x00
-    sta 0x2116  ; VRAM address low
-    sta 0x2117  ; VRAM address high
-    sta 0x211A  ; initial mode 7 setting register
-    sta 0x211B  ; mode 7 matrix parameter A register (low)
-    lda #0x01
-    sta 0x211B  ; mode 7 matrix parameter A register (high)
-    lda #0x00
-    sta 0x211C  ; mode 7 matrix parameter B register (low)
-    sta 0x211C  ; mode 7 matrix parameter B register (high)
-    sta 0x211D  ; mode 7 matrix parameter C register (low)
-    sta 0x211D  ; mode 7 matrix parameter C register (high)
-    sta 0x211E  ; mode 7 matrix parameter D register (low)
-    lda #0x01
-    sta 0x211E  ; mode 7 matrix parameter D register (high)
-    lda #0x00
-    sta 0x211F  ; mode 7 center position X register (low)
-    sta 0x211F  ; mode 7 center position X register (high)
-    sta 0x2120  ; mode 7 center position Y register (low)
-    sta 0x2120  ; mode 7 center position Y register (high)
-    sta 0x2121  ; color number register (0x00-0xff)
-    sta 0x2123  ; bg1 & bg2 window mask setting register
-    sta 0x2124  ; bg3 & bg4 window mask setting register
-    sta 0x2125  ; obj & color window mask setting register
-    sta 0x2126  ; window 1 left position register
-    sta 0x2127  ; window 2 left position register
-    sta 0x2128  ; window 3 left position register
-    sta 0x2129  ; window 4 left position register
-    sta 0x212A  ; bg1, bg2, bg3, bg4 window logic register
-    sta 0x212B  ; obj, color window logic register (or, and, xor, xnor)
-    lda #0x01
-    sta 0x212C  ; main screen designation (planes, sprites enable)
-    lda #0x00
-    sta 0x212D  ; sub screen designation
-    sta 0x212E  ; window mask for main screen
-    sta 0x212F  ; window mask for sub screen
-    lda #0x30
-    sta 0x2130  ; color addition & screen addition init setting
-    lda #0x00
-    sta 0x2131  ; add/sub sub designation for screen, sprite, color
-    lda #0xE0
-    sta 0x2132  ; color data for addition/subtraction
-    stz 0x2133  ; screen setting (interlace x,y/enable SFX data)
-    stz 0x4200  ; disable v-blank, interrupt, joypad register
-    lda #0xFF
-    sta 0x4201  ; programmable I/O port
-    lda #0x00
-    sta 0x4202  ; multiplicand A
-    sta 0x4203  ; multiplier B
-    sta 0x4204  ; multiplier C
-    sta 0x4205  ; multiplicand C
-    sta 0x4206  ; divisor B
-    sta 0x4207  ; horizontal count timer
-    sta 0x4208  ; horizontal count timer MSB
-    sta 0x4209  ; vertical count timer
-    sta 0x420A  ; vertical count timer MSB
-    sta 0x420B  ; general DMA enable (bits 0-7)
-    sta 0x420C  ; horizontal DMA (HDMA) enable (bits 0-7)
-    sta 0x420D  ; access cycle designation (slow/fast rom)
-    rts
+    enable_gamepad:
+    """Enable auto-joypad-read + vblank NMI by writing 1 to $4200."""
+        pha
+        lda #0x01
+        sta 0x4200
+        pla
+        rts
+
+
+    disable_gamepad:
+    """Disable auto-joypad-read and all NMI/IRQ sources ($4200 := 0)."""
+        stz 0x4200
+        rts
+
+
+    initialize_snes:
+    """
+    Reset PPU, DMA, and CPU registers to a known idle state at boot.
+    8-bit M/X  ; clears OBJ/BG registers, neutralises mode-7 matrix to identity,
+    disables NMI/HDMA/joypad, leaves screen forced-blank.
+    """
+        sep #0x30  ; make X, Y, A all 8-bits
+        lda #0x80  ; screen off, no brightness
+        sta 0x2100  ; brightness & screen enable register
+        lda #0x00
+        sta 0x2101  ; sprite register (size & address in VRAM)
+        sta 0x2102  ; sprite registers (address of sprite memory [OAM])
+        sta 0x2103  ; sprite registers (address of sprite memory [OAM])
+        sta 0x2105  ; graphic mode register
+        sta 0x2106  ; mosaic register
+        sta 0x2107  ; plane 0 map VRAM location
+        sta 0x2108  ; plane 1 map VRAM location
+        sta 0x2109  ; plane 2 map VRAM location
+        sta 0x210A  ; plane 3 map VRAM location
+        sta 0x210B  ; plane 0 & 1 Tile data location
+        sta 0x210C  ; plane 2 & 3 Tile data location
+        sta 0x210D  ; plane 0 scroll x (first 8 bits)
+        sta 0x210D  ; plane 0 scroll x (last 3 bits)
+        sta 0x210E  ; plane 0 scroll y (first 8 bits)
+        sta 0x210E  ; plane 0 scroll y (last 3 bits)
+        sta 0x210F  ; plane 1 scroll x (first 8 bits)
+        sta 0x210F  ; plane 1 scroll x (last 3 bits)
+        sta 0x2110  ; plane 1 scroll y (first 8 bits)
+        sta 0x2110  ; plane 1 scroll y (last 3 bits)
+        sta 0x2111  ; plane 2 scroll x (first 8 bits)
+        sta 0x2111  ; plane 2 scroll x (last 3 bits)
+        sta 0x2112  ; plane 2 scroll y (first 8 bits)
+        sta 0x2112  ; plane 2 scroll y (last 3 bits)
+        sta 0x2113  ; plane 3 scroll x (first 8 bits)
+        sta 0x2113  ; plane 3 scroll x (last 3 bits)
+        sta 0x2114  ; plane 3 scroll y (first 8 bits)
+        sta 0x2114  ; plane 3 scroll y (last 3 bits)
+        lda #0x80  ; increase VRAM address after writing to 0x2119
+        sta 0x2115  ; VRAM address increment register
+        lda #0x00
+        sta 0x2116  ; VRAM address low
+        sta 0x2117  ; VRAM address high
+        sta 0x211A  ; initial mode 7 setting register
+        sta 0x211B  ; mode 7 matrix parameter A register (low)
+        lda #0x01
+        sta 0x211B  ; mode 7 matrix parameter A register (high)
+        lda #0x00
+        sta 0x211C  ; mode 7 matrix parameter B register (low)
+        sta 0x211C  ; mode 7 matrix parameter B register (high)
+        sta 0x211D  ; mode 7 matrix parameter C register (low)
+        sta 0x211D  ; mode 7 matrix parameter C register (high)
+        sta 0x211E  ; mode 7 matrix parameter D register (low)
+        lda #0x01
+        sta 0x211E  ; mode 7 matrix parameter D register (high)
+        lda #0x00
+        sta 0x211F  ; mode 7 center position X register (low)
+        sta 0x211F  ; mode 7 center position X register (high)
+        sta 0x2120  ; mode 7 center position Y register (low)
+        sta 0x2120  ; mode 7 center position Y register (high)
+        sta 0x2121  ; color number register (0x00-0xff)
+        sta 0x2123  ; bg1 & bg2 window mask setting register
+        sta 0x2124  ; bg3 & bg4 window mask setting register
+        sta 0x2125  ; obj & color window mask setting register
+        sta 0x2126  ; window 1 left position register
+        sta 0x2127  ; window 2 left position register
+        sta 0x2128  ; window 3 left position register
+        sta 0x2129  ; window 4 left position register
+        sta 0x212A  ; bg1, bg2, bg3, bg4 window logic register
+        sta 0x212B  ; obj, color window logic register (or, and, xor, xnor)
+        lda #0x01
+        sta 0x212C  ; main screen designation (planes, sprites enable)
+        lda #0x00
+        sta 0x212D  ; sub screen designation
+        sta 0x212E  ; window mask for main screen
+        sta 0x212F  ; window mask for sub screen
+        lda #0x30
+        sta 0x2130  ; color addition & screen addition init setting
+        lda #0x00
+        sta 0x2131  ; add/sub sub designation for screen, sprite, color
+        lda #0xE0
+        sta 0x2132  ; color data for addition/subtraction
+        stz 0x2133  ; screen setting (interlace x,y/enable SFX data)
+        stz 0x4200  ; disable v-blank, interrupt, joypad register
+        lda #0xFF
+        sta 0x4201  ; programmable I/O port
+        lda #0x00
+        sta 0x4202  ; multiplicand A
+        sta 0x4203  ; multiplier B
+        sta 0x4204  ; multiplier C
+        sta 0x4205  ; multiplicand C
+        sta 0x4206  ; divisor B
+        sta 0x4207  ; horizontal count timer
+        sta 0x4208  ; horizontal count timer MSB
+        sta 0x4209  ; vertical count timer
+        sta 0x420A  ; vertical count timer MSB
+        sta 0x420B  ; general DMA enable (bits 0-7)
+        sta 0x420C  ; horizontal DMA (HDMA) enable (bits 0-7)
+        sta 0x420D  ; access cycle designation (slow/fast rom)
+        rts
+}
diff --git a/src/menus/in_game_text.s b/src/menus/in_game_text.s
index 707ea2a..91c6034 100644
--- a/src/menus/in_game_text.s
+++ b/src/menus/in_game_text.s
@@ -1,357 +1,361 @@
 """French translated text data for the in-game menus."""
 .include "src/ingame/macros.i"
+.include "../bank20.i"
 .table "text/ff4_menus.tbl"
 .extern lookup_dakuten
 
-.scope in_game_menu {
-    """Main pause-menu strings."""
-cant_fight:
-    .text "KO"
-    .db 0
-menu:
-; window
-    menu_window(23, 0, 7, 17)
-; position
-    .dw 0x0070
-    .text "Objets"
-    .db 0x01
-    .dw 0x00F0
-    .text "Magie"
-    .db 0x01
-    .dw 0x0170
-    .text "Equiper"
-    .db 0x01
-    .dw 0x01F0
-    .text "Statut"
-    .db 0x01
-    .dw 0x0270
-    .text "Placer"
-    .db 0x01
-    .dw 0x02F0
-    .text "Changer"
-    .db 0x01
-    .dw 0x0370
-    .text "Options"
-    .db 0x01
-    .dw 0x03F0
-    .text "Sauver"
-    .db 0
-gils:
-    .text "Gils"
-    .db 0
-time:
-    .text "Temps"
-    .db 0
-}
 
-.scope items_menu {
-    """Items submenu strings."""
-items_menu_right:
-    menu_window(22, 0, 7, 3)
-item:
-    menu_window(0, 0, 7, 3)
-    .dw 0x0044
-    .text "Objets"
-    .db 0
-notuse:
-    .dw 0x0052
-    .text "Impossible à utiliser."
-    .db 0
-}
+.alloc in_game_text_block in bank20_reloc {
+    .scope in_game_menu {
+        """Main pause-menu strings."""
+    cant_fight:
+        .text "KO"
+        .db 0
+    menu:
+    ; window
+        menu_window(23, 0, 7, 17)
+    ; position
+        .dw 0x0070
+        .text "Objets"
+        .db 0x01
+        .dw 0x00F0
+        .text "Magie"
+        .db 0x01
+        .dw 0x0170
+        .text "Equiper"
+        .db 0x01
+        .dw 0x01F0
+        .text "Statut"
+        .db 0x01
+        .dw 0x0270
+        .text "Placer"
+        .db 0x01
+        .dw 0x02F0
+        .text "Changer"
+        .db 0x01
+        .dw 0x0370
+        .text "Options"
+        .db 0x01
+        .dw 0x03F0
+        .text "Sauver"
+        .db 0
+    gils:
+        .text "Gils"
+        .db 0
+    time:
+        .text "Temps"
+        .db 0
+    }
 
+    .scope items_menu {
+        """Items submenu strings."""
+    items_menu_right:
+        menu_window(22, 0, 7, 3)
+    item:
+        menu_window(0, 0, 7, 3)
+        .dw 0x0044
+        .text "Objets"
+        .db 0
+    notuse:
+        .dw 0x0052
+        .text "Impossible à utiliser."
+        .db 0
+    }
 
-.scope spells {
-    """Spell-list headers."""
-white:
-    move_to(24, 3)
-    .text "Magie"
-    .db 0
-black:
-    move_to(24, 5)
-    .text "Rituel"
-    .db 0
-summon:
-    move_to(24, 7)
-    .text "Chimere"
-    .db 0
-ninja:
-    move_to(24, 5)
-    .text "Ninja "
-    .db 0
-kokan:
-    move_to(24, 3)
-    .text "Echange"
-    .db 0
-mp_needed:
-    .text "Coût PM"
-    .db 0
-}
 
-.scope status {
-    """Status screen labels."""
-status:
-    .dw 0x01F0
-    .text "Statut"
-    .db 0
-exp_for_next_level:
-    .dw 0x0260
-    .text "Niveau suivant"
-    .db 0
-char_stats:
-    .dw 0x0114 - 0x80
-    .text "Niveau"
-    .db 1
-    .dw 0x01A0
-    .text "Expérience"
-    .db 1
-    .dw 0x0206
-    .text "PV"
-    .db 1
-    .dw 0x0286
-    .text "PM"
-    .db 1
-    .dw 0x0344
-    .text "Talents"
-    .db 1
-    .dw 0x03C2
-    .text "Vigueur"
-    .db 1
-    .dw 0x0442
-    .text "Agilité"
-    .db 1
-    .dw 0x04C2
-    .text "Vitesse"
-    .db 1
-    .dw 0x0542
-    .text "Esprit"
-    .db 1
-    .dw 0x05C2
-    .text "Volonté"
-    .db 1
-;att/def/mag:
-    .dw 0x035A
-    .text "Attaque"
-    .db 1
-    .dw 0x03DA
-    .text "Attaque%"
-    .db 1
-    .dw 0x045A
-    .text "Défense"
-    .db 1
-    .dw 0x04DA
-    .text "Défense%"
-    .db 1
-    .dw 0x055A
-    .text "Déf Mag"
-    .db 1
-    .dw 0x05DA
-    .text "Déf Mag%"
-    .db 0
-}
+    .scope spells {
+        """Spell-list headers."""
+    white:
+        move_to(24, 3)
+        .text "Magie"
+        .db 0
+    black:
+        move_to(24, 5)
+        .text "Rituel"
+        .db 0
+    summon:
+        move_to(24, 7)
+        .text "Chimere"
+        .db 0
+    ninja:
+        move_to(24, 5)
+        .text "Ninja "
+        .db 0
+    kokan:
+        move_to(24, 3)
+        .text "Echange"
+        .db 0
+    mp_needed:
+        .text "Coût PM"
+        .db 0
+    }
 
-.scope options {
-    """Options screen text."""
-title:
-    .dw 0x0096
-    .text "Options"
-    .db 0
-config:
-    .dw 0x0102, 0x141C
-    .dw 0x0144
-    .text "Mode Combat"
-    .db 0x01
-    .dw 0x015E
-    .text "Actif  Pause"
-    .db 0x01
-    .dw 0x01C4
-    .text "Vit. Combat"
-    .db 0x01
-    .dw 0x021E
-    .text "Vite   Lent"
-    .db 0x01
-    .dw 0x0244
-    .text "Vit. Texte"
-    .db 0x01
-    .dw 0x02C4
-    .text "Audio"
-    .db 0x01
-    .dw 0x02DE
-    .text "Stéréo Mono"
-    .db 0x01
-    .dw 0x0344
-    .text "Contrôle"
-    .db 0x01
-    .dw 0x035E
-    .text "Normal Perso."
-    .db 0x01
-    .dw 0x03DE
-    .text "Seul   Multiple"
-    .db 0x01
-    .dw 0x0444
-    .text "Curseur"
-    .db 0x01
-    .dw 0x045E
-    .text "Reset  Mémoire"
-    .db 0x01
-    .dw 0x04C4
-    .text "Couleur"
-    .db 0
-controls:
-    move_to(5, 1)
-    .text "Contrôles Personalisés"
-    .db 0x01
-    .dw 0x0204
-    .text "Action"
-    .db 0x01
-    .dw 0x0284
-    .text "Annuler"
-    .db 0x01
-    .dw 0x0304
-    .text "Menu"
-    .db 0x01
-    .dw 0x0384
-    .text "Left Button"
-    .db 0x01
-    .dw 0x0404
-    .text "Start"
-    .db 1
-    .dw 0x0484
-    .text "Fin"
-    .db 0
-}
+    .scope status {
+        """Status screen labels."""
+    status:
+        .dw 0x01F0
+        .text "Statut"
+        .db 0
+    exp_for_next_level:
+        .dw 0x0260
+        .text "Niveau suivant"
+        .db 0
+    char_stats:
+        .dw 0x0114 - 0x80
+        .text "Niveau"
+        .db 1
+        .dw 0x01A0
+        .text "Expérience"
+        .db 1
+        .dw 0x0206
+        .text "PV"
+        .db 1
+        .dw 0x0286
+        .text "PM"
+        .db 1
+        .dw 0x0344
+        .text "Talents"
+        .db 1
+        .dw 0x03C2
+        .text "Vigueur"
+        .db 1
+        .dw 0x0442
+        .text "Agilité"
+        .db 1
+        .dw 0x04C2
+        .text "Vitesse"
+        .db 1
+        .dw 0x0542
+        .text "Esprit"
+        .db 1
+        .dw 0x05C2
+        .text "Volonté"
+        .db 1
+    ;att/def/mag:
+        .dw 0x035A
+        .text "Attaque"
+        .db 1
+        .dw 0x03DA
+        .text "Attaque%"
+        .db 1
+        .dw 0x045A
+        .text "Défense"
+        .db 1
+        .dw 0x04DA
+        .text "Défense%"
+        .db 1
+        .dw 0x055A
+        .text "Déf Mag"
+        .db 1
+        .dw 0x05DA
+        .text "Déf Mag%"
+        .db 0
+    }
 
-.scope equip {
-    """Equipment menu slot labels."""
-menu:
-    _text_y = 1
-    menu_window(0, 0, 30, 11)
-    move_to(13, 0 + _text_y)
-    .text "M. Droite"
-    .db 0x01
-    move_to(13, 2 + _text_y)
-    .text "M. Gauche"
-    .db 0x01
-    move_to(13, 4 + _text_y)
-    .text "Tête"
-    .db 0x01
-    move_to(13, 6 + _text_y)
-    .text "Corps"
-    .db 0x01
-    move_to(13, 8 + _text_y)
-    .text "Mains"
-    .db 0
-}
+    .scope options {
+        """Options screen text."""
+    title:
+        .dw 0x0096
+        .text "Options"
+        .db 0
+    config:
+        .dw 0x0102, 0x141C
+        .dw 0x0144
+        .text "Mode Combat"
+        .db 0x01
+        .dw 0x015E
+        .text "Actif  Pause"
+        .db 0x01
+        .dw 0x01C4
+        .text "Vit. Combat"
+        .db 0x01
+        .dw 0x021E
+        .text "Vite   Lent"
+        .db 0x01
+        .dw 0x0244
+        .text "Vit. Texte"
+        .db 0x01
+        .dw 0x02C4
+        .text "Audio"
+        .db 0x01
+        .dw 0x02DE
+        .text "Stéréo Mono"
+        .db 0x01
+        .dw 0x0344
+        .text "Contrôle"
+        .db 0x01
+        .dw 0x035E
+        .text "Normal Perso."
+        .db 0x01
+        .dw 0x03DE
+        .text "Seul   Multiple"
+        .db 0x01
+        .dw 0x0444
+        .text "Curseur"
+        .db 0x01
+        .dw 0x045E
+        .text "Reset  Mémoire"
+        .db 0x01
+        .dw 0x04C4
+        .text "Couleur"
+        .db 0
+    controls:
+        move_to(5, 1)
+        .text "Contrôles Personalisés"
+        .db 0x01
+        .dw 0x0204
+        .text "Action"
+        .db 0x01
+        .dw 0x0284
+        .text "Annuler"
+        .db 0x01
+        .dw 0x0304
+        .text "Menu"
+        .db 0x01
+        .dw 0x0384
+        .text "Left Button"
+        .db 0x01
+        .dw 0x0404
+        .text "Start"
+        .db 1
+        .dw 0x0484
+        .text "Fin"
+        .db 0
+    }
 
+    .scope equip {
+        """Equipment menu slot labels."""
+    menu:
+        _text_y = 1
+        menu_window(0, 0, 30, 11)
+        move_to(13, 0 + _text_y)
+        .text "M. Droite"
+        .db 0x01
+        move_to(13, 2 + _text_y)
+        .text "M. Gauche"
+        .db 0x01
+        move_to(13, 4 + _text_y)
+        .text "Tête"
+        .db 0x01
+        move_to(13, 6 + _text_y)
+        .text "Corps"
+        .db 0x01
+        move_to(13, 8 + _text_y)
+        .text "Mains"
+        .db 0
+    }
 
-.scope dextrality {
-    """Handedness labels."""
-hands:
-string_0:
-; ぶきよう
-    .text "String 0"
-    .db 0
-string_1:
-; ひだりきき
-    .text "Gaucher"
-    .db 0
-string_2:
-; みぎきき
-    .text "Droiter"
-    .db 0
-string_3:
-; りょうきき
-    .text "Ambidextre"
-    .db 0
-}
 
-.scope messages {
-    """Generic prompt strings."""
-use_on_whom:
-    move_to(10, 1)
-    .text "Utiliser sur qui ?"
-    .db 0
-cantuse:
-    move_to(10, 1)
-    .text "Cet objet ne peut être utilisé ici."
-    .db 0
-cant_use_magic:
-    menu_window_move_text(7, 12, 14, 1)
-    .text "Ne peut utiliser la magie"
-    .db 0
-}
+    .scope dextrality {
+        """Handedness labels."""
+    hands:
+    string_0:
+    ; ぶきよう
+        .text "String 0"
+        .db 0
+    string_1:
+    ; ひだりきき
+        .text "Gaucher"
+        .db 0
+    string_2:
+    ; みぎきき
+        .text "Droiter"
+        .db 0
+    string_3:
+    ; りょうきき
+        .text "Ambidextre"
+        .db 0
+    }
 
+    .scope messages {
+        """Generic prompt strings."""
+    use_on_whom:
+        move_to(10, 1)
+        .text "Utiliser sur qui ?"
+        .db 0
+    cantuse:
+        move_to(10, 1)
+        .text "Cet objet ne peut être utilisé ici."
+        .db 0
+    cant_use_magic:
+        menu_window_move_text(7, 12, 14, 1)
+        .text "Ne peut utiliser la magie"
+        .db 0
+    }
 
-.scope use_spell {
-    """Spell-cast prompt strings."""
-mp_cost:
-    move_to(1, 4)
-    .text "Requis"
-    .db 0
-; use_on_whom:
-; move_to(1, 10)
-; .text 'Sur qui ?'
-; .db 0
-}
 
-.scope treasure {
-    """Treasure-chest UI text + windows."""
-choice_window:
-"""patch them in place"""
-    menu_window(8, 0, 22, 2)
-items_window:
-    menu_window(0, 3, 30, 10)
-header_window:
-    menu_window(0, 0, 6, 2)
-    move_to(1, 1)
-    .text "Butin"
-    .db 1
-exit:
-    move_to(24, 1)
-    .text "Quitter"
-    .db 1
-take_all:
-    move_to(10, 1)
-    .text "Tout prendre"
-    .db 0
-exchange:
-    move_to(10, 1)
-    .text "Échanger    "
-    .db 0
-key_items_left_warning:
-    menu_window(5, 10, 19, 4)
-    move_to(6, 11)
-    .text "  Il reste des    "
-    .db 1
-    move_to(6, 13)
-    .text "objets importants."
-    .db 0
-}
+    .scope use_spell {
+        """Spell-cast prompt strings."""
+    mp_cost:
+        move_to(1, 4)
+        .text "Requis"
+        .db 0
+    ; use_on_whom:
+    ; move_to(1, 10)
+    ; .text 'Sur qui ?'
+    ; .db 0
+    }
+
+    .scope treasure {
+        """Treasure-chest UI text + windows."""
+    choice_window:
+    """patch them in place"""
+        menu_window(8, 0, 22, 2)
+    items_window:
+        menu_window(0, 3, 30, 10)
+    header_window:
+        menu_window(0, 0, 6, 2)
+        move_to(1, 1)
+        .text "Butin"
+        .db 1
+    exit:
+        move_to(24, 1)
+        .text "Quitter"
+        .db 1
+    take_all:
+        move_to(10, 1)
+        .text "Tout prendre"
+        .db 0
+    exchange:
+        move_to(10, 1)
+        .text "Échanger    "
+        .db 0
+    key_items_left_warning:
+        menu_window(5, 10, 19, 4)
+        move_to(6, 11)
+        .text "  Il reste des    "
+        .db 1
+        move_to(6, 13)
+        .text "objets importants."
+        .db 0
+    }
 
 
-copy_text_with_dakuten_far:
-"""Far-callable text copier with dakuten composite lookup."""
-{
-    phb
-    phk
-    plb
-    rep #0x20
-    txa
-    clc
-    adc 0x29
-    tax
-    sep #0x20
-_loop:
-    lda.w 0x0000, y
-    beq _exit
-    jsr.l lookup_dakuten
-    sta 0x7e0000, x
-    xba
-    sta 0x7e0040, x
-    inx
-    inx
-    iny
-    bra _loop
-_exit:
-    plb
-    rtl
+    copy_text_with_dakuten_far:
+    """Far-callable text copier with dakuten composite lookup."""
+    {
+        phb
+        phk
+        plb
+        rep #0x20
+        txa
+        clc
+        adc 0x29
+        tax
+        sep #0x20
+    _loop:
+        lda.w 0x0000, y
+        beq _exit
+        jsr.l lookup_dakuten
+        sta 0x7e0000, x
+        xba
+        sta 0x7e0040, x
+        inx
+        inx
+        iny
+        bra _loop
+    _exit:
+        plb
+        rtl
+    }
 }
diff --git a/src/menus/start_screen_text.s b/src/menus/start_screen_text.s
index b5af106..f325e4b 100644
--- a/src/menus/start_screen_text.s
+++ b/src/menus/start_screen_text.s
@@ -1,58 +1,62 @@
 """French translated text data for the start screen / save selection."""
 .include "src/ingame/macros.i"
+.include "../bank20.i"
 
 .table "text/ff4_menus.tbl"
-.scope newgame {
-    """Title-screen + save-slot strings."""
-new_game:
-    move_to(1, 1)
-    .text "Nouvelle partie"
-    .db 0
-    .if DEBUG {
-build_number:
-    VERSION := 'v1.0.0a0'
-    .text "${BUILD_DATE} ${VERSION}"
-    .db 0
+
+.alloc start_screen_text_block in bank20_reloc {
+    .scope newgame {
+        """Title-screen + save-slot strings."""
+    new_game:
+        move_to(1, 1)
+        .text "Nouvelle partie"
+        .db 0
+        .if DEBUG {
+    build_number:
+        VERSION := 'v1.0.0a0'
+        .text "${BUILD_DATE} ${VERSION}"
+        .db 0
+        }
+    time_load_save:
+        .dw 0x046E + 2
+        .text "Temps"
+        .db 0
+    gils_load_game:
+        .text "Gils"
+        .db 0
+    save:
+        .text "Partie"
+        .db 0
+    load_this_save:
+        .dw 0x006E + 2
+        .text "Cette"
+        .db 1
+        .dw 0x00EE + 2
+        .text "partie?"
+        .db 0
+    yes_no:
+        .dw 0x0172
+        .text "Oui"
+        .db 1
+        .dw 0x01F2
+        .text "Non"
+        .db 0
+    empty_save:
+        .dw 0x0090 + 2
+        .text "VIDE"
+        .db 0
+    saves:
+        move_to(1, 1)
+        .text "Sauvegardes"
+        .db 0
+    save_completed:
+        move_to(8, 11)
+        .text "Sauvegarde terminée"
+        .db 0
+    did_not_save:
+        move_to(1, 1)
+        .text "Annulation "
+    ; extra space at the end to clear the previous title.
+        .db 0
     }
-time_load_save:
-    .dw 0x046E + 2
-    .text "Temps"
-    .db 0
-gils_load_game:
-    .text "Gils"
-    .db 0
-save:
-    .text "Partie"
-    .db 0
-load_this_save:
-    .dw 0x006E + 2
-    .text "Cette"
-    .db 1
-    .dw 0x00EE + 2
-    .text "partie?"
-    .db 0
-yes_no:
-    .dw 0x0172
-    .text "Oui"
-    .db 1
-    .dw 0x01F2
-    .text "Non"
-    .db 0
-empty_save:
-    .dw 0x0090 + 2
-    .text "VIDE"
-    .db 0
-saves:
-    move_to(1, 1)
-    .text "Sauvegardes"
-    .db 0
-save_completed:
-    move_to(8, 11)
-    .text "Sauvegarde terminée"
-    .db 0
-did_not_save:
-    move_to(1, 1)
-    .text "Annulation "
-; extra space at the end to clear the previous title.
-    .db 0
 }
diff --git a/src/menus/system_menus_text.s b/src/menus/system_menus_text.s
index cf77f9d..c17c9fb 100644
--- a/src/menus/system_menus_text.s
+++ b/src/menus/system_menus_text.s
@@ -1,128 +1,133 @@
 """Relocated implementations of the original system-menu text routines."""
-.extern assets_classes_ptr
-
-.macro _bank_switch() {
-    cpy.w #0x8000
-    bmi _moved_text
-    lda.b #0x01
-; this assumes we shoud load data from bank 1
-    pha
-    bra _jump_to_original
-_moved_text:
-    pha
-    rep #0x20
-    tya
-    adc.w #0x8000
-    tay
-    sep #0x20
-    pla
-    phk
-_jump_to_original:
-"""this assumes that the text lives in the same bank as this routine."""
-    plb
-}
-
-.macro _bank_switch_with_jump(jump_to) {
-    _bank_switch()
-    jmp.l jump_to
-}
+.include "../bank20.i"
 
+.extern assets_classes_ptr
 
-display_text_in_menus:
-"""Trampoline into the original menu text-display routine at $01:830B, preserving DBR/D/X via stack."""
-{
-    phb
-    phd
-    phx
-    ldx.w #0x100
-    phx
-    pld
 
-    _bank_switch_with_jump(0x01830B)
+.alloc system_menus_text_block in bank20_reloc {
+    .macro _bank_switch() {
+        cpy.w #0x8000
+        bmi _moved_text
+        lda.b #0x01
+    ; this assumes we shoud load data from bank 1
+        pha
+        bra _jump_to_original
+    _moved_text:
+        pha
+        rep #0x20
+        tya
+        adc.w #0x8000
+        tay
+        sep #0x20
+        pla
+        phk
+    _jump_to_original:
+    """this assumes that the text lives in the same bank as this routine."""
+        plb
+    }
+
+    .macro _bank_switch_with_jump(jump_to) {
+        _bank_switch()
+        jmp.l jump_to
+    }
+
+
+    display_text_in_menus:
+    """Trampoline into the original menu text-display routine at $01:830B, preserving DBR/D/X via stack."""
+    {
+        phb
+        phd
+        phx
+        ldx.w #0x100
+        phx
+        pld
+
+        _bank_switch_with_jump(0x01830B)
+    }
+
+    load_text_with_destination_in_x:
+    """Adjust X by the offset in $29 and trampoline into the original load-text routine at $01:8318."""
+
+        phb
+        phd
+        phx
+        phx
+        ldx.w #0x0100
+        phx
+        pld
+        plx
+
+        rep #0x20
+
+        txa
+        clc
+        adc 0x29
+        tax
+        sep #0x20
+
+        _bank_switch_with_jump(0x018318)
+
+
+    display_window_with_text:
+    """Trampoline into the original window+text drawer at $01:80DD."""
+        phy
+        phb
+        _bank_switch_with_jump(0x0180DD)
+
+    display_time:
+    """Trampoline into the original play-time display routine at $01:879D."""
+        phb
+        _bank_switch()
+        rep #0x20
+        jmp.l 0x01879D
+
+    disable_save:
+    """
+    Overwrite the 'Sauver' menu label with $24 (space tile) glyphs to grey out the save option, then continue at
+    $01:8947.
+    """
+        lda.b #0x24
+        sta 0xCA31  ; S
+        sta 0xCA33  ; a
+        sta 0xCA35  ; u
+        sta 0xCA37  ; v
+        sta 0xCA39  ; e
+        sta 0xCA3B  ; r
+        jmp.l 0x018947
+
+    load_classes_pointer:
+    """Resolve a class-name pointer from `assets_classes_ptr[A*2]` into A (24-bit)."""
+        phx
+        rep #0x20
+
+        and.w #0x00FF
+        asl
+        tax
+
+        lda.l assets_classes_ptr, x
+        sep #0x20
+
+        plx
+        rtl
+
+    load_dextrelity_pointer:
+    """
+    Read the high two bits of (DP),$60 to pick a dexterity-text pointer from the table at $01:E2D9  ; returns
+    24-bit pointer in Y.
+    """
+        rep #0x20
+        lda (0x60)
+        and.w #0x00c0
+        lsr
+        lsr
+        lsr
+        lsr
+        lsr
+        phx
+        tax
+        lda.l 0x01e2d9, x
+        plx
+        tay
+        sep #0x20
+        rtl
 }
-
-load_text_with_destination_in_x:
-"""Adjust X by the offset in $29 and trampoline into the original load-text routine at $01:8318."""
-
-    phb
-    phd
-    phx
-    phx
-    ldx.w #0x0100
-    phx
-    pld
-    plx
-
-    rep #0x20
-
-    txa
-    clc
-    adc 0x29
-    tax
-    sep #0x20
-
-    _bank_switch_with_jump(0x018318)
-
-
-display_window_with_text:
-"""Trampoline into the original window+text drawer at $01:80DD."""
-    phy
-    phb
-    _bank_switch_with_jump(0x0180DD)
-
-display_time:
-"""Trampoline into the original play-time display routine at $01:879D."""
-    phb
-    _bank_switch()
-    rep #0x20
-    jmp.l 0x01879D
-
-disable_save:
-"""
-Overwrite the 'Sauver' menu label with $24 (space tile) glyphs to grey out the save option, then continue at
-$01:8947.
-"""
-    lda.b #0x24
-    sta 0xCA31  ; S
-    sta 0xCA33  ; a
-    sta 0xCA35  ; u
-    sta 0xCA37  ; v
-    sta 0xCA39  ; e
-    sta 0xCA3B  ; r
-    jmp.l 0x018947
-
-load_classes_pointer:
-"""Resolve a class-name pointer from `assets_classes_ptr[A*2]` into A (24-bit)."""
-    phx
-    rep #0x20
-
-    and.w #0x00FF
-    asl
-    tax
-
-    lda.l assets_classes_ptr, x
-    sep #0x20
-
-    plx
-    rtl
-
-load_dextrelity_pointer:
-"""
-Read the high two bits of (DP),$60 to pick a dexterity-text pointer from the table at $01:E2D9  ; returns
-24-bit pointer in Y.
-"""
-    rep #0x20
-    lda (0x60)
-    and.w #0x00c0
-    lsr
-    lsr
-    lsr
-    lsr
-    lsr
-    phx
-    tax
-    lda.l 0x01e2d9, x
-    plx
-    tay
-    sep #0x20
-    rtl
diff --git a/src/menus/tools_shop_text.s b/src/menus/tools_shop_text.s
index 3946ac4..7fad818 100644
--- a/src/menus/tools_shop_text.s
+++ b/src/menus/tools_shop_text.s
@@ -1,78 +1,92 @@
 """French translated text data for the tools/weapon/armor shop UI."""
 .include "src/ingame/macros.i"
+.include "../bank20.i"
 
 .table "text/ff4_menus.tbl"
-.scope shops {
-    """Shop UI strings."""
-gils:
-    move_to(27, 6)
-    .text "Gils"
-    .db 0
-welcome_and_actions:
-    .dw 0x0054 - 2
-    .text "Puis-je vous aider ?"
-;.text 'いらっしゃい! どんなごようけんで?'
-    .db 1
-    .dw 0x0148 - 4
-    .text "Achat Vente Sortir"
-    .db 0
-quantity:
-""".text 'かう   うる   でる'"""
-    .dw 0x0052
-    .text "Que désirez vous ?  "
-    .db 1
-    .dw 0x0144
-    .text "Quantité"
-    .db 1
-    .dw 0x0146 + 14 * 2
-    .text "1"
-    .db 0
-thank_you_window:
-    menu_window(5, 10, 11, 2)
-thank_you_text:
-    move_to(6, 11)
-    .text "Merci !"
-    .db 0
-inventory_full:
-    menu_window(1, 10, 20, 4)
-    move_to(2, 11)
-    .text "L\'inventaire est"
-    .db 1
-    move_to(2, 13)
-    .text "plein."
-    .db 0
-not_enough_gils:
-    menu_window(5, 10, 16, 4)
-    move_to(6, 11)
-    .text "Vous n\'avez pas"
-    .db 1
-    move_to(6, 13)
-    .text "de Gils."
-    .db 0
-sell_window:
-    menu_window(8, 10, 14, 13)
-    move_to(13, 13)
-    .text " Unités"
-    .db 1
-    move_to(19, 15)
-    .text "Gils"
-    .db 1
-    move_to(10, 17)
-    .text "Êtes-vous"
-    .db 1
-    move_to(11, 19)
-    .text "d\'accord ?"
-    .db 1
-    move_to(12, 21)
-    .text "Oui Non"
-    .db 0
-weapons_title:
-    .text "Armes"
-    .db 0
-armor_title:
-    .text "Armures"
-    .db 0
-items_title:
-    .text "Objets"
-    .db 0
+
+.alloc tools_shop_text_block in bank20_reloc {
+    .scope shops {
+        """Shop UI strings."""
+    gils:
+        move_to(27, 6)
+        .text "Gils"
+        .db 0
+    puis_je_vous_aider:
+    """Owner welcome greeting  ; rendered through the small-VWF description region."""
+        .dw 0x0054 - 2
+        .text "Puis-je vous aider ?"
+    ;.text 'いらっしゃい! どんなごようけんで?'
+        .db 0
+    welcome_and_actions:
+        .dw 0x0148 - 4
+        .text "Achat Vente Sortir"
+        .db 0
+    que_desirez_vous:
+    """Owner welcome prompt  ; rendered through the small-VWF description region."""
+        .dw 0x0052
+        .text "Que désirez vous ?  "
+        .db 0
+    quantity:
+    """.text 'かう   うる   でる'"""
+        .dw 0x0144
+        .text "Quantité"
+        .db 1
+        .dw 0x0146 + 14 * 2
+        .text "1"
+        .db 0
+    thank_you_window:
+        menu_window(5, 10, 11, 2)
+    ; Trailing empty-text block keeps the vanilla `$82FB` (draw window + text)
+    ; happy when called via shop_thanks_text_hook ; the actual "Merci !" copy
+    ; is rendered separately through the small-VWF description region.
+        .dw 0x0000
+        .db 0
+    merci:
+    """Owner thank-you message  ; rendered through the small-VWF description region."""
+        move_to(6, 11)
+        .text "Merci !"
+        .db 0
+    inventory_full:
+        menu_window(1, 10, 20, 4)
+        move_to(2, 11)
+        .text "L\'inventaire est"
+        .db 1
+        move_to(2, 13)
+        .text "plein."
+        .db 0
+    not_enough_gils:
+        menu_window(5, 10, 16, 4)
+        move_to(6, 11)
+        .text "Vous n\'avez pas"
+        .db 1
+        move_to(6, 13)
+        .text "de Gils."
+        .db 0
+    sell_window:
+        menu_window(8, 10, 14, 13)
+        move_to(13, 13)
+        .text " Unités"
+        .db 1
+        move_to(19, 15)
+        .text "Gils"
+        .db 1
+        move_to(10, 17)
+        .text "Êtes-vous"
+        .db 1
+        move_to(11, 19)
+        .text "d\'accord ?"
+        .db 1
+        move_to(12, 21)
+        .text "Oui Non"
+        .db 0
+    weapons_title:
+        .text "Armes"
+        .db 0
+    armor_title:
+        .text "Armures"
+        .db 0
+    items_title:
+        .text "Objets"
+        .db 0
+    }
 }
diff --git a/src/menus/tools_shop_text_en.s b/src/menus/tools_shop_text_en.s
index 8b19481..5f0ac37 100644
--- a/src/menus/tools_shop_text_en.s
+++ b/src/menus/tools_shop_text_en.s
@@ -7,19 +7,23 @@ gils:
     move_to(27, 6)
     .text "GP"
     .db 0
-welcome_and_actions:
+puis_je_vous_aider:
+"""Owner welcome greeting  ; rendered through the small-VWF description region."""
     .dw 0x0054 - 2
     .text "Welcome!"
 ;.text 'いらっしゃい! どんなごようけんで?'
-    .db 1
+    .db 0
+welcome_and_actions:
     .dw 0x0148 - 4
     .text "Buy Sell Exit"
     .db 0
-quantity:
-""".text 'かう   うる   でる'"""
+que_desirez_vous:
+"""Owner welcome prompt  ; rendered through the small-VWF description region."""
     .dw 0x0052
     .text "What do you want?   "
-    .db 1
+    .db 0
+quantity:
+""".text 'かう   うる   でる'"""
     .dw 0x0144
     .text "Quantity"
     .db 1
@@ -28,7 +32,10 @@ quantity:
     .db 0
 thank_you_window:
     menu_window(5, 10, 11, 2)
-thank_you_text:
+    .dw 0x0000
+    .db 0
+merci:
+"""Owner thank-you message  ; rendered through the small-VWF description region."""
     move_to(6, 11)
     .text "Thank you!"
     .db 0
diff --git a/src/small_vwf/init.s b/src/small_vwf/init.s
index 9d8ccb3..fe396a6 100644
--- a/src/small_vwf/init.s
+++ b/src/small_vwf/init.s
@@ -3,14 +3,18 @@ Small (8x8) VWF init routine: prepares the menu tile buffer, runs the renderer i
 RTL entry points for cross-bank callers.
 """
 .include "src/vwf.i"
+.include "../bank20.i"
 .include "src/libmz.i"
 
 .import "assets"
 .import "libmz"
-.if BATTLE_ENABLED {
-    .extern battle_render
-    .extern battle_render.clear_buffer
-}
 
-.include "src/small_vwf/render.s"
-.include "src/small_vwf/item_description.s"
+.alloc small_vwf_init_block in bank20_reloc {
+    .if BATTLE_ENABLED {
+        .extern battle_render
+        .extern battle_render.clear_buffer
+    }
+
+    .include "src/small_vwf/render.s"
+    .include "src/small_vwf/item_description.s"
+}
diff --git a/src/small_vwf/item_description.s b/src/small_vwf/item_description.s
index 4b7cef1..92e090d 100644
--- a/src/small_vwf/item_description.s
+++ b/src/small_vwf/item_description.s
@@ -1,4 +1,7 @@
 """Small-VWF item-description renderer."""
+.include "src/vwf_state.i"
+.include "src/items.i"
+
 .scope items_description {
     """Small-VWF item-description renderer entry-points."""
 draw_trampoline:
@@ -49,7 +52,62 @@ draw:
     plx
     sep #0x20
 draw_string:
+; Region config: description CHR lives in tile_id range $180..$1FF
+; (allocator base $80, with attr-byte OR $01 -> effective tile_id
+; $180 ; VRAM byte $5800..$5FF0). Disjoint from the field-items
+; window at $100..$169 -> VRAM $5000..$56A0 ; the two regions used
+; to fight for VRAM $5000 (both DMAs targeted there) and the last
+; render stomped the other.
+;
+;   Region map (BG3 CHR at VRAM $4000)
+;     $00..$FF   static menu font ($4000..$4FF0)
+;     $100..$169 field-menu items ($5000..$56A0)
+;     $180..$1FF item description  ($5800..$5FF0)
+    rep #0x20
+    lda.w #0x0080
+    sta.l VWF_CONFIG_BASE + VwfConfig.tile_id_base
+    sep #0x20
+    lda.b #0x80
+    sta.l VWF_CONFIG_BASE + VwfConfig.slot_budget
+    lda.b #0x01
+    sta.l VWF_CONFIG_BASE + VwfConfig.flags
+; Description flush descriptor uses the SECONDARY slot so the
+; description region ($5800..$5FF0) gets its own DMA independent of
+; the primary descriptor that items_menu_vwf rewrites every per-slot
+; call. The allocator base is $80 (`init_with_tile_id_wide(#$0080)`)
+; with the $01 attr-byte OR turning it into effective tile_id $180
+; for tilemap entries ; the actual CHR bytes live at buffer offset
+; $80 * 16 = $800 (NOT $1800).
+;
+;   src offset = $80 * 16             = $0800 buffer bytes
+;   vram dest  = ($5000 + $800) / 2   = $2C00 word
+;   size       = $80 tiles * 16       = $0800 bytes
+;
+; DIRTY_B is set at the tail of draw_string ; display_char also sets
+; primary DIRTY which the field-items renderer covers in its own
+; flush, so no clearing is needed here.
+    rep #0x20
+    lda.w #0x0800
+    sta.l VWF_CHR_SRC_OFFSET_B
+    lda.w #0x2C00
+    sta.l VWF_CHR_VRAM_WORD_B
+    lda.w #0x0800
+    sta.l VWF_CHR_BYTE_COUNT_B
+    sep #0x20
+; Preserve Y across render.init: the per-region CHR clear loop
+; tays the budget word count and lands Y=$0000 on exit, which
+; collapsed _char_loop into an immediate _char_loop_exit (the loop
+; reads `lda.w $0000,y` -> $20:0000 = $00 terminator) and the
+; description never rendered a single glyph.
+    phy
     jsr.w render.init
+    rep #0x20
+    lda.w #0x0080
+    jsr.w render_allocator.init_with_tile_id_wide
+    sep #0x20  ; _char_loop reads bytes one at a time ; M=16 here
+; would lda two source bytes per char and soft-lock
+; in wait_for_vblank.
+    ply
 _char_loop:
     lda.w 0x0000, y
     beq _char_loop_exit
@@ -67,6 +125,10 @@ _char_loop:
 _char_loop_exit:
     jsr.w _transfer_item_description
     jsr.w render.deinit
+; Description finished writing to buffer ; raise the secondary
+; dirty flag so the NMI flush DMAs $5800..$5FF0 next vblank.
+    lda.b #0x01
+    sta.l VWF_CHR_DIRTY_B
     plx
     pld
     plb
@@ -99,8 +161,10 @@ _reset_render:
     bra _char_loop
 _transfer_item_description:
     jsr.w wait_for_vblank
-; we could reduce the transfer size to the size of the string
-    dma_transfer_to_vram_call(render.buffer_ptr, 0x5000 >> 1, render.buffer_size, 0x1801)
-; use of the menu engine built in dma transfer in NMI
+; DMA description's CHR slice only: SRAM $703800 (= buffer + $80*16
+; for tile_id_base $80) to VRAM word $2C00 (= byte $5800), $800
+; bytes covering 128 tile_ids. Keeps the upload disjoint from the
+; field-items DMA at byte $5000 / word $2800.
+    dma_transfer_to_vram_call(render.buffer_ptr + 0x800, 0x5800 >> 1, 0x800, 0x1801)
     rts
 }
diff --git a/src/small_vwf/render.s b/src/small_vwf/render.s
index 73448f8..4287e7a 100644
--- a/src/small_vwf/render.s
+++ b/src/small_vwf/render.s
@@ -1,4 +1,6 @@
 """Small (8x8) menu VWF renderer."""
+.include "src/vwf_state.i"
+
 VARS_BUFFER = 0x710000
 
 .macro initialize(var) {
@@ -31,7 +33,10 @@ VARS_BUFFER = 0x710000
 }
 
 .scope _vram_copy {
-    buffer = 0x704000
+; Moved from $704000 to $705000 so the CHR buffer at $703000 can
+; grow to $1000 bytes (cover tile_ids $00..$FF) without trampling
+; the vram-save staging.
+    buffer = 0x705000
 save_dialog_vram_far:
     jsr.l 0x14fd0f
 ; original save
@@ -107,16 +112,75 @@ _transfer_to_vram:
 }
 
 .scope render_allocator {
-    """Tile-id allocator for the small-VWF."""
+    """
+    Tile-id allocator for the small-VWF. Stored as a 16-bit word at
+    $702F00..$702F01 so callers can reach the full 10-bit BG3
+    tile_id range (0..0x3FF) instead of being capped at 8 bits.
+    All accessors operate in 16-bit M and write both bytes  ; callers
+    that only need the low byte can still do an 8-bit lda.l at the
+    same address.
+    """
     allocated_tile_id = 0x702F00
+    slot_limit_low = 0x702F02
+"""
+Inventory clamp. When the allocator's low byte reaches this value,
+`increment` freezes it  ; the blitter happily keeps reading the same
+tile_id and overwrites the slot's last tile instead of bleeding into
+the next slot's CHR. Set to 0xFF (no clamp) by every non-inventory
+init. Inventory slot init writes slot_base + K-1 here.
+"""
+
+
 init_with_tile_id:
-"""tile_id in A"""
+"""
+8-bit caller convention: A.lo = tile_id (0..0xFF), M = 8.
+Zero-extends to 16-bit so the high byte at $702F01 is clean.
+tilemap_write_no_inc adds the hardcoded +0x100 bit shift. 16-bit
+callers needing tile_ids outside the 0x100-0x1FF window should use
+init_with_tile_id_wide. Also resets slot_limit_low to 0xFF so the
+clamp is a no-op for non-inventory regions.
+"""
+
+
     sta.l allocated_tile_id
+    pha
+    lda.b #0xFF
+    sta.l slot_limit_low
+    pla
+    php
+    rep #0x20
+    pha
+    lda.l allocated_tile_id
+    and.w #0x00ff
+    sta.l allocated_tile_id
+    pla
+    plp
+    rts
+init_with_tile_id_wide:
+"""
+16-bit caller convention: A = full tile_id (0..0x3FF), M = 16.
+Stores both bytes verbatim. Use this for inventory / future regions
+that go past tile_id 0xFF. Also resets slot_limit_low to 0xFF so the
+clamp is a no-op until the inventory slot init opts in.
+"""
+
+
+    sta.l allocated_tile_id
+    pha
+    sep #0x20
+    lda.b #0xFF
+    sta.l slot_limit_low
+    rep #0x20
+    pla
     rts
 init:
     pha
-    lda.b #0x00
+    rep #0x20
+    lda.w #0x0000
     sta.l allocated_tile_id
+    sep #0x20
+    lda.b #0xFF
+    sta.l slot_limit_low
     pla
     rts
     .if BATTLE_ENABLED {
@@ -127,14 +191,32 @@ init_battle_far:
     rtl
     }
 increment:
+"""
+Advance the 16-bit allocator by 1. Preserves M / A state.
+Clamps the low byte at `slot_limit_low` so an inventory slot whose
+name overflows K tiles keeps overwriting its last tile instead of
+bleeding into the next slot's tile_id range. Non-inventory regions
+set slot_limit_low = 0xFF so the clamp never fires.
+"""
+
+
+    php
+    sep #0x20
     pha
     lda.l allocated_tile_id
+    cmp.l slot_limit_low
+    bcs _inc_clamped
+    rep #0x20
+    lda.l allocated_tile_id
     inc
-    and #0xff
     sta.l allocated_tile_id
+    sep #0x20
+_inc_clamped:
     pla
+    plp
     rts
 get:
+"""Return current 16-bit allocator value in A (caller sets M)."""
     lda.l allocated_tile_id
     rts
 }
@@ -149,14 +231,24 @@ get:
     prev_char = counter + 2
     current_char = prev_char + 1
     tilemap_offset = 0x1d
-    buffer_ptr = 0x703000
-    buffer_size = 0x300
+    buffer_ptr = VWF_CHR_BUFFER
+    buffer_size = VWF_CHR_BUFFER_SIZE
     last_drawn_text_ptr = buffer_ptr + buffer_size + 2
 init:
-"""font_ptr = assets_menu_font_dat  ; moved to direct use of assets_menu_font_dat"""
-; Initialize the renderer
-; clear a chunk of ram
-; resets variables
+"""
+font_ptr = assets_menu_font_dat  ; moved to direct use of assets_menu_font_dat.
+
+Resets the per-render scratch (bits_left_on_tile / temp / counter)
+and zeros the allocator. Clears ONLY the CHR slice owned by the
+current region (`VwfConfig.tile_id_base * 16` for
+`VwfConfig.slot_budget * 16` bytes), mirror of the per-slot clear
+in battle's `init_inventory_for_current_slot`. Wiping the whole
+buffer here clobbered other regions' CHR mid-frame, so callers
+populate config first and `init` confines its $FF/$00 fill to
+their slot.
+"""
+
+
     .if ENABLE_KERNING_MENU {
     stz.b prev_char
     }
@@ -170,22 +262,347 @@ _brk_init_bits:
     initialize(temp)
     stz.b temp
     pha
-    lda #0x00
-    phx
-    ldx.w #0
-_clear_loop:
+; --- Per-region CHR clear from VwfConfig ---
+    php
+    rep #0x30
+    lda.l VWF_CONFIG_BASE + VwfConfig.tile_id_base
+    and.w #0x01FF  ; 9-bit tile_id_base
+    asl
+    asl
+    asl
+    asl  ; * 16
+    clc
+    adc.w #VWF_CHR_BUFFER & 0xFFFF
+    tax
+    lda.l VWF_CONFIG_BASE + VwfConfig.slot_budget
+    and.w #0x00FF
+    asl
+    asl
+    asl  ; budget * 8 word-pairs
+    tay
+    beq _init_skip_clear
+    sep #0x20
+
+_init_clear_loop:
     lda.b #0xFF
-    sta.l buffer_ptr, x
+    sta.l VWF_CHR_BUFFER & 0xFF0000, x
+    inx
     lda.b #0x00
-    sta.l buffer_ptr + 1, x
+    sta.l VWF_CHR_BUFFER & 0xFF0000, x
     inx
+    dey
+    bne _init_clear_loop
+
+_init_skip_clear:
+    plp
+    pla  ; balances the `pha` after `stz.b temp` ; without this the
+; `rts` popped the saved A byte as the high byte of the
+; return address and item_description.draw_trampoline ran
+; off into a BRK opcode on every menu open.
+    initialize(counter)
+    rts
+flush_chr_to_vram:
+"""
+NMI-callable VWF CHR flush. Gates on `VWF_CHR_DIRTY`  ; if set,
+reads `VwfConfig.chr_src_offset` / `chr_vram_word` / `chr_byte_count`
+from the config struct and DMAs the slice from `VWF_CHR_BUFFER +
+chr_src_offset` to VRAM word `chr_vram_word`. Clears the dirty
+byte on the way out.
+
+Designed to hang off the field NMI hook  ; running in vblank means
+the DMA does not tear visible scanlines and we drop the explicit
+`wait_for_vblank` the synchronous tail-of-render flush needed.
+Battle-side equivalent of the partial CHR DMA in
+`messages_vwf.dma_transfer`.
+"""
+
+
+    {
+    php
+    sep #0x20
+    rep #0x10
+    lda.l VWF_CHR_DIRTY
+    beq _flush_skip_a
+    lda.b #0x00
+    sta.l VWF_CHR_DIRTY
+; CRITICAL: clear HDMAEN bit 6 before retargeting ch6 for a one-shot
+; DMA. The treasure-popup rolling buffer also arms ch6 (BG3VOFS HDMA)
+; via $1BAE ; if we rewrite $4360-$4365 with the VWF flush descriptor
+; without disabling HDMA bit 6 first, the per-scanline HDMA keeps
+; running off the rewritten registers and writes VWF_CHR_BUFFER bytes
+; to VMDATAL on every visible scanline for the rest of the frame.
+; That trashes VRAM, kills BG3 scroll modulation, and made the
+; treasure-popup band display garbage stride after the first frame.
+; Push HDMAEN on the stack so we restore the prior mask exactly
+; (treasure ch6 + drops ch4 + anything else vanilla armed).
+    lda.l 0x00420C
+    pha
+    and.b #0xBF
+    sta.l 0x00420C
+; Channel 6: free per the FF4 DMA audit (vanilla btlgfx + menu
+; never touch $4360..$436F ; ch7 already carries battle's
+; `_sram_dma_transfer_7`, small_vwf's libmz transfers and the
+; field NMI's tilemap upload). Source = VWF_CHR_BUFFER +
+; VWF_CHR_FLUSH_OFFSET (engine constant, identical for battle +
+; field). Dest VRAM word + size come from VwfConfig so each caller
+; targets its own CHR slot without forking the upload path.
+    lda.b #0x01  ; DMAP: word transfer (2 byte regs, alt low/high)
+    sta.l 0x004360
+    lda.b #0x18  ; BBAD: $2118 VMDATAL
+    sta.l 0x004361
+    rep #0x20
+    lda.w #( VWF_CHR_BUFFER + VWF_CHR_FLUSH_OFFSET ) & 0xFFFF
+    sta.l 0x004362  ; A1T low+mid
+    sep #0x20
+    lda.b #( VWF_CHR_BUFFER + VWF_CHR_FLUSH_OFFSET ) >> 16
+    sta.l 0x004364  ; A1B source bank
+    rep #0x20
+    lda.l VWF_CONFIG_BASE + VwfConfig.chr_vram_word
+    sta.l 0x002116  ; VMADD
+    lda.l VWF_CONFIG_BASE + VwfConfig.chr_byte_count
+    sta.l 0x004365  ; DAS
+    sep #0x20
+    lda.b #0x80
+    sta.l 0x002115  ; VMAIN: increment on $2119, +1 word
+    lda.b #0x40
+    sta.l 0x00420B  ; MDMAEN ch6
+; Restore prior HDMAEN bits (treasure ch6, drops ch4, etc) ; one-shot
+; DMA on ch6 has completed by now and HDMA can resume reading the
+; treasure HDMA table for per-scanline BG3VOFS on the next frame.
+    pla
+    sta.l 0x00420C
+
+_flush_skip_a:
+; --- Secondary descriptor flush (region 1B, drops in the treasure
+; popup). Shares DMA channel 6 with the primary flush ; runs after
+; the primary descriptor's DMA finishes so we never re-arm a busy
+; channel. Each descriptor covers a disjoint slice of VRAM so two
+; sequential DMAs land at distinct dest addresses without conflict.
+; Secondary source offset is its OWN word so drops can DMA only its
+; CHR slice ; primary keeps starting at VWF_CHR_FLUSH_OFFSET.
+;
+; Guard against savestates captured before the secondary descriptor
+; existed : their SRAM at $7070C4..$7070CA carries random bytes, so
+; before firing the DMA we whitelist VWF_CHR_VRAM_WORD_B against the
+; two known arming targets ($2B70 drops, $2C00 description) and
+; reject anything else as uninitialised. Stale SRAM almost certainly
+; will not match either constant, so the secondary DMA stays dormant
+; until items_menu_vwf or item_description explicitly arms it. Stops
+; the save-selection sprite CHR from being trashed by random garbage
+; on the very first NMI after a stale load.
+    sep #0x20
+    lda.l VWF_CHR_DIRTY_B
+    beq _flush_skip_b
+    lda.b #0x00
+    sta.l VWF_CHR_DIRTY_B
+    rep #0x20
+    lda.l VWF_CHR_VRAM_WORD_B
+    cmp.w #0x2B70
+    beq _vram_word_b_ok
+    cmp.w #0x2C00
+    bne _flush_skip_b_late
+
+_vram_word_b_ok:
+    sep #0x20
+; Same HDMAEN-bit-6 save/restore dance as the primary flush ; ch6 is
+; shared with treasure BG3VOFS HDMA and must not stay armed while
+; we hijack its registers for the secondary VRAM blit.
+    lda.l 0x00420C
+    pha
+    and.b #0xBF
+    sta.l 0x00420C
+    lda.b #0x01
+    sta.l 0x004360
+    lda.b #0x18
+    sta.l 0x004361
+    rep #0x20
+    lda.l VWF_CHR_SRC_OFFSET_B
+    clc
+    adc.w #VWF_CHR_BUFFER & 0xFFFF
+    sta.l 0x004362
+    sep #0x20
+    lda.b #VWF_CHR_BUFFER >> 16
+    sta.l 0x004364
+    rep #0x20
+    lda.l VWF_CHR_VRAM_WORD_B
+    sta.l 0x002116
+    lda.l VWF_CHR_BYTE_COUNT_B
+    sta.l 0x004365
+    sep #0x20
+    lda.b #0x80
+    sta.l 0x002115
+    lda.b #0x40
+    sta.l 0x00420B
+    pla
+    sta.l 0x00420C
+
+_flush_skip_b_late:
+    sep #0x20
+
+_flush_skip_b:
+    plp
+    rts
+    }
+render_with_config:
+"""
+Config-driven VWF entry: reads `VwfConfig` at `VWF_CONFIG_BASE`,
+sets up the allocator + render state from it, then walks
+VWF_TEXT_BUFFER through `draw_text_buffer`. Single call site
+replaces the bespoke init / display_char loops that the field-
+items helper and the battle inventory text walker each carry.
+
+VwfConfig fields consumed (matches `src/vwf_state.i`):
+  tile_id_base   first tile_id this slot owns
+  slot_budget    +K-1 clamp  ; $FF = no clamp
+  tilemap_base   absolute WRAM byte index of the destination
+                 tilemap row  ; display_char advances tilemap_offset
+                 (DP $1D) by 2 per blit
+
+`font_ptr` / `kerning_ptr` are reserved in the struct but
+display_char still hardcodes `assets_menu_font_dat` for this
+phase  ; swapping the font fetch to indirect-through-config is the
+next refactor and lets battle keep its own font without forking
+the engine.
+
+M=8, X=16 on entry. Stack-balanced, RTS.
+"""
+
+
+    {
+    php
+    sep #0x20
+    rep #0x10
+; Reset per-glyph carry state so each slot's first char starts fresh.
+; prev_char carries over from the previous slot's last char ; without
+; a reset the new slot's first kerning lookup pairs against an
+; unrelated glyph and shifts the allocator several tiles. Same goes
+; for bits_left_on_tile : if a prior render ended with the value
+; pinned by the $09 clamp at _end (kerning overshoot path), the new
+; slot would start with bits_left=0 and burn its first tile.
+; Symptom before this reset : treasure-inventory slot 0 rendered as
+; "le d'or" (Aigu chopped) right after drops finished rendering.
+    .if ENABLE_KERNING_MENU {
+    stz.b prev_char
+    }
+    lda.b #0x08
+    sta.b bits_left_on_tile
+; Allocator base + clamp via the 16-bit `init_with_tile_id_wide`
+; entry. tile_id_base is a 9-bit value (0..$1FF) so we need the
+; full word ; the 8-bit init zero-extends and caps at $FF which
+; would have wrapped field-menu slots 6..10 (base $C0 + 10*10 =
+; $124) back into the menu font CHR range.
+    rep #0x20
+    lda.l VWF_CONFIG_BASE + VwfConfig.tile_id_base
+    jsr.w render_allocator.init_with_tile_id_wide
+; Belt + braces: re-stash allocated_tile_id verbatim so any
+; interleaved render.init can not zero it on us before display_char
+; reads it back.
+    lda.l VWF_CONFIG_BASE + VwfConfig.tile_id_base
+    sta.l render_allocator.allocated_tile_id
+    sep #0x20
+    lda.l VWF_CONFIG_BASE + VwfConfig.tile_id_base
+    clc
+    adc.l VWF_CONFIG_BASE + VwfConfig.slot_budget
+    sec
+    sbc.b #0x01
+    sta.l render_allocator.slot_limit_low
+; --- Clear this slot's CHR slice to $FF $00 (blank 2bpp / 4bpp
+; tile) so the new blit does not pile pixels onto the previous
+; tenant. Mirror of `messages_vwf.init_inventory_for_current_slot`
+; in `src/battle/message.s` ; clear loop spans
+; (slot_budget * 16) bytes at VWF_CHR_BUFFER + tile_id_base * 16.
+    php
+    rep #0x30
+    lda.l VWF_CONFIG_BASE + VwfConfig.tile_id_base
+    and.w #0x01FF  ; 9-bit tile_id_base
+    asl
+    asl
+    asl
+    asl  ; * 16
+    clc
+    adc.w #VWF_CHR_BUFFER & 0xFFFF
+    tax
+    lda.l VWF_CONFIG_BASE + VwfConfig.slot_budget
+    and.w #0x00FF
+    asl
+    asl
+    asl  ; words to clear = budget * 8 (16 bytes/tile, written as words)
+    tay
+    sep #0x20
+
+_chr_clear_loop:
+    lda.b #0xFF
+    sta.l VWF_CHR_BUFFER & 0xFF0000, x
+    inx
+    lda.b #0x00
+    sta.l VWF_CHR_BUFFER & 0xFF0000, x
+    inx
+    dey
+    bne _chr_clear_loop
+    plp
+; Render scratch state.
+    lda.b #0x08
+    sta.b bits_left_on_tile
+    stz.b temp
+    stz.b counter
+; tilemap_offset = config.tilemap_base (16-bit).
+    rep #0x20
+    lda.l VWF_CONFIG_BASE + VwfConfig.tilemap_base
+    sta.b tilemap_offset
+    sep #0x20
+    jsr.w draw_text_buffer
+; Tell the NMI flush hook this slot needs a VRAM upload. Mirror of
+; `battle_render.dma_dirty_slots` ; engine-side so every consumer
+; (field items, item descriptions, treasure list, ...) signals dirty
+; without knowing about VRAM addresses.
+    lda.b #0x01
+    sta.l VWF_CHR_DIRTY
+    plp
+    rts
+    }
+
+draw_text_buffer:
+"""
+Unified entry: walk a null-terminated string at VWF_TEXT_BUFFER
+and blit each byte via display_char.
+
+Caller responsibilities (read from VwfConfig at VWF_CONFIG_BASE in
+the next phase  ; for now the field-items helper hand-sets these
+directly):
+  - render_allocator.allocated_tile_id    set via init_with_tile_id
+  - render_allocator.slot_limit_low       set per slot budget
+  - render.bits_left_on_tile              set to 8
+  - render.temp, render.counter           cleared
+  - render.tilemap_offset (= DP $1D)      absolute WRAM byte index
+                                          of the bottom tilemap row
+                                          start  ; display_char auto-
+                                          increments by 2 per blit
+  - X, Y                                  free for caller use
+
+Each iteration reads the next byte from `VWF_TEXT_BUFFER + Y` (Y
+caller-zeroed on entry), exits on $00. display_char writes both
+the CHR slice and the tile_id at tilemap_offset, and increments
+the allocator (clamped at slot_limit_low).
+"""
+
+
+    {
+    phx
+    ldx.w #0x0000
+
+_dtb_loop:
+    lda.l VWF_TEXT_BUFFER, x
+    beq _dtb_done
     inx
-    cpx.w #buffer_size + 2
-    bne _clear_loop
+    phx
+    jsr.w display_char
+    plx
+    bra _dtb_loop
+
+_dtb_done:
     plx
-    pla
-    initialize(counter)
     rts
+    }
 deinit:
     {
     _restore_long(buffer_ptr)
@@ -216,14 +633,16 @@ make_pointers:
     lda.w #0x0000
     sep #0x20
 
-    lda.l render_allocator.allocated_tile_id
-
-    sta.b oldtilepos
+; Read full 16-bit allocated_tile_id so 9-bit tile_ids ($100+) keep
+; their high bit in oldtilepos. The previous 8-bit read masked bit
+; 8 and field-menu VWF blits (tile_id_base $100) landed at SRAM
+; offset 0, stomping the static font region.
     rep #0x20
+    lda.l render_allocator.allocated_tile_id
     asl
     asl
     asl
-    asl
+    asl  ; tile_id * 16, 16-bit result
     sta.b oldtilepos
     tay
     sep #0x20
@@ -443,7 +862,20 @@ _no_adjustment:
 
     pla
 _end:
+; Clamp temp to 0..8 before writing bits_left_on_tile : kerning
+; adjustments can push the value past 8 (e.g. when a kerning entry
+; subtracts more than the current tile has left), and the downstream
+; dispatch at $97FC indexes a 9-entry jump table by bits_left * 2.
+; Out-of-range values land off the table end and execute random
+; bytes (BRK $00 on the held-DOWN field-items scroll). Treat any
+; value > 8 as "no bits left in this tile" so the next loop pass
+; allocates a fresh tile_id instead of crashing.
     lda.b temp
+    cmp.b #0x09
+    bcc _bits_left_in_range
+    lda.b #0x00
+
+_bits_left_in_range:
     sta.b bits_left_on_tile
 _overflow:
     pha
@@ -571,12 +1003,26 @@ small_vwf_kerning_binary_ext:
     rtl
     }
 tilemap_write_no_inc:
+"""
+Write `allocated_tile_id` low byte to `(tilemap_offset)` and OR
+`VwfConfig.flags` into the next (palette/attr) byte. The flag
+mask used to be hard-coded `$01` (the 9-bit tile_id bit for the
+512-tile dialog window) but the field-menu items live in 2bpp
+BG3 where bit 0 of the attr byte is tile_id bit 8  ; an always-on
+OR pushed all our tile_ids into the +256 half of CHR and the
+high inventory slots wrapped past $FF in unpredictable ways.
+Letting the caller pick the OR value via the config struct keeps
+dialog (set `flags = $01`) working without forcing the menu items
+to honour a bit they do not own.
+"""
+
+
     _base_addr = 0x7e0000
     lda.l render_allocator.allocated_tile_id
     ldx.b tilemap_offset
     sta.l _base_addr, x
     lda.l _base_addr + 1, x
-    ora.b #0x01
+    ora.l VWF_CONFIG_BASE + VwfConfig.flags
     sta.l _base_addr + 1, x
     rts
 tilemap_write:
diff --git a/src/sram_management.s b/src/sram_management.s
index 7c3d8ec..33adcca 100644
--- a/src/sram_management.s
+++ b/src/sram_management.s
@@ -18,9 +18,54 @@ ROM-header SRAM-size byte at $00:FFD8 (clamped to 7 → 128KB max).
     jsr.l 0x15C9AA
     lda.b #1
     jsr.w _clear_ram
-
+    jsr.w _zero_vwf_description_region
     rtl
 
+_zero_vwf_description_region:
+"""
+Belt-and-suspenders boot init : zero the menu VWF VRAM union
+
+$5000-$5FF0 ($1000 bytes from VRAM word $2800). Covers :
+  Region 1   $5000-$56A0  (field / treasure item-name CHR)
+  Region 1B  $56E0-$5AA0  (drops item-name CHR, treasure popup)
+  Region D   $5800-$5FF0  (item description CHR)
+
+Without this, the menu save/restore round-trips whatever leftover
+bytes happened to be at these VRAM addresses at the very first
+SaveDlgGfx, which then persist forever through the SRAM save buffer
+as glyph residue in the VRAM viewer (overworld never reads these
+slices). Overworld re-uploads its own map CHR on every map transition,
+so this one-time zero does not erase live overworld data  ; it only
+breaks the carry-forward of stale boot-VRAM.
+
+SRAM $70:0000+ was just zeroed by `_clear_ram` above, so it works as
+a fixed-source zero-fill DMA.
+"""
+
+
+    php
+    sep #0x20
+    lda.b #0x80
+    sta.l 0x002100  ; PPU force blank
+    lda.b #0x80
+    sta.l 0x002115  ; VMAIN : increment on $2119, +1 word
+    rep #0x20
+    lda.w #0x2800  ; VMADD : VRAM word $2800 (= byte $5000)
+    sta.l 0x002116
+    lda.w #0x1811  ; $4300 = $11 (mode 1, fixed src) | $4301 = $18 (VMDATAL)
+    sta.l 0x004300
+    lda.w #0x0000  ; A1T : SRAM byte $00 (freshly zeroed by _clear_ram)
+    sta.l 0x004302
+    lda.w #0x1000  ; DAS : $1000 byte transfer = $5000..$5FFF
+    sta.l 0x004305
+    sep #0x20
+    lda.b #0x70  ; A1B : SRAM bank $70
+    sta.l 0x004304
+    lda.b #0x01
+    sta.l 0x00420B  ; MDMAEN ch0
+    plp
+    rts
+
 _clear_ram:
 {
 ; Get RAM size from ROM header
diff --git a/src/vwf.s b/src/vwf.s
index a13a894..656f209 100644
--- a/src/vwf.s
+++ b/src/vwf.s
@@ -3,6 +3,7 @@ Dialog/text VWF rendering engine: control-code parser (`parse`), per-glyph blit
 tracking (`TILEPOS`/`BITSLEFT`), button-glyph + ending-symbol drawing, and the gil-window tilemaps.
 """
 .include "src/definitions.s"
+.include "bank20.i"
 .include "src/libmz.i"
 .table "text/ff4_menus.tbl"
 .import "libmz"
@@ -11,1190 +12,1193 @@ tracking (`TILEPOS`/`BITSLEFT`), button-glyph + ending-symbol drawing, and the g
 .import "assets"
 .import "kerning"
 
-wait_for_action_button:
-"""RTL trampoline that polls the gamepad until any action button is pressed (delegates to original `_waitpad`)."""
-    jsr.w _waitpad
-    rtl
 
-.if ENABLE_BUTTON_DISPLAY {
-_get_action_button_id:
-    lda.l 0x0016A9
-    bne _custom_mapping
-    lda.b #0x00
-    xba
-    lda.b #0x00
-    rts
-_custom_mapping:
-    lda.b #0x00
-    xba
-    lda.l 0x001A37  ; action button id location
-    rts
-}
+.alloc vwf_block in bank20_reloc {
+    wait_for_action_button:
+    """RTL trampoline that polls the gamepad until any action button is pressed (delegates to original `_waitpad`)."""
+        jsr.w _waitpad
+        rtl
 
-.macro _make_color(r, g, b) {
-    red := ( r >> 3 ) & 0b11111
-    blue := ( b >> 3 ) & 0xb11111
-    green := ( g >> 3 ) & 0xb11111
-    .dw ( blue << 10 ) + ( green << 5 ) + red
-}
+    .if ENABLE_BUTTON_DISPLAY {
+    _get_action_button_id:
+        lda.l 0x0016A9
+        bne _custom_mapping
+        lda.b #0x00
+        xba
+        lda.b #0x00
+        rts
+    _custom_mapping:
+        lda.b #0x00
+        xba
+        lda.l 0x001A37  ; action button id location
+        rts
+    }
 
-.macro hex_color(color) {
-    """Emit a 15-bit BGR SNES palette word from a 24-bit literal."""
-    _make_color(color & 0xff, ( color >> 8 ) & 0xff, ( color >> 16 ))
-}
+    .macro _make_color(r, g, b) {
+        red := ( r >> 3 ) & 0b11111
+        blue := ( b >> 3 ) & 0xb11111
+        green := ( g >> 3 ) & 0xb11111
+        .dw ( blue << 10 ) + ( green << 5 ) + red
+    }
 
+    .macro hex_color(color) {
+        """Emit a 15-bit BGR SNES palette word from a 24-bit literal."""
+        _make_color(color & 0xff, ( color >> 8 ) & 0xff, ( color >> 16 ))
+    }
 
-_button_colors:
-    ; 0
-    ; A Button
-    _make_color(0xeb, 0x1a, 0x1d)
-    _make_color(120, 10, 12)
-    ; 1
-    ; B Button
-    _make_color(0xfe, 0xce, 0x15)
-    _make_color(136, 108, 0)
-    ; 2
-    ; X Button
-    _make_color(0x07, 0x49, 0xb4)
-    _make_color(0, 53, 144)
-    ; 3
-    ; Y Button
-    _make_color(0x00, 0x8d, 0x45)
-    _make_color(0, 70, 34)
-    ; 4
-    ; Select Buttons
-    .dw 0b0011110111101111
-    .dw 0b0001110011100111
-
-
-update_palette:
-"""
-Refresh menu/dialog palette entries from the configured colour at $16AA.
-With ENABLE_BUTTON_DISPLAY, also pulls the action-button id and writes
-the matching colour + shadow pair from `_button_colors` into $0CEF/$0CF1.
-"""
-    ldx.w 0x16AA
-    stx.w 0x0CDD
-    stx.w 0x0CE5
-.if ENABLE_BUTTON_DISPLAY {
-    stx.w 0x0CED
-; black for the shadow but it could be a darker version of the color
-    ldx.w #0x0000
-    stx.w 0x0CEF
-    pha
-    jsr.w _get_action_button_id
-
-    cmp #0x04
-    bcc _proceed
-    ldx.w #5
-_proceed:
-    asl
-    asl
-    tax
-
-    rep #0x20
-    lda.l _button_colors, x  ; color
-    sta.w 0x0CF1
-    lda.l _button_colors + 2, x  ; shadow color
-    sta.w 0x0CEF
-    sep #0x20
-    lda #0x00
-    xba
-    pla
-}
-rtl
-
-window_palette:
-"""4 dialog/menu sub-palettes (4 colours each, BGR15)."""
-    ; Menu color palette I probably need to add one color
-    ; Palette 1, normal text
-    .db 0x00, 0x00
-    .db 0x00, 0x40
-    .db 0xCE, 0x39
-    .db 0xFF, 0x7F
-
-; Palette 2, greyed out text patched replaced the grey with black
-    .db 0x00, 0x00
-    .db 0x00, 0x40
-    .db 0x00, 0x00
-    .db 0xff, 0x7f
-
-; Palette  3, yellow
-    .db 0x00, 0x00
-    .db 0x00, 0x40
-    .db 0x80, 0x02
-    .db 0x7F, 0x03
-
-; Palette 4, red
-    .db 0x00, 0x00
-    .db 0x00, 0x40
-    .db 0xFF, 0x40
-    .db 0x7F, 0x2E
-
-;current_pos = 0x3d
-;current_text_pointer = 0x0772
-; 0x07 used for loop counter
-
-; ******************
-; ** Declarations **
-; ******************
-    vram_tile_set_pointer = 0x6800
-    vram_tile_map_pointer = 0x2C00
-    WRAM = field_vwf.tile_buffer
-
-    WRAMPTR = 0x2108
-
-vwfinit:
-"""
-One-shot dialog VWF initialisation: clear the WRAM scratch area,
-DMA the VWF tile buffer + the original font tileset into VRAM,
-and point BG3 at the new $6000 tile-set base.
-"""
-    jsr.w _clr  ; on efface un peu de Wram
-
-    jsr.w wait_for_vblank
-    dma_transfer_to_vram_call(WRAM, vram_tile_set_pointer, 0x0690, 0x1801)
-    jsr.w wait_for_vblank
-    dma_transfer_to_vram_call(WRAM, vram_tile_set_pointer + 0x348, 0x0690, 0x1801)
-
-; copy the old font tileset
-    jsr.w wait_for_vblank
-    dma_transfer_to_vram_call(0x0AF000, 0x6000, 0x800, 0x1801)
-    jsr.w wait_for_vblank
-    dma_transfer_to_vram_call(0x0AF000 + 0x800, 0x6000 + 0x400, 0x800, 0x1801)
-    jsr.w wait_for_vblank
-
-; Sets the BG3 vram pointer to 0x6000
-    lda 0x210C
-    and #0xF0
-    clc
-    adc #0x06
-    sta 0x210C
 
+    _button_colors:
+        ; 0
+        ; A Button
+        _make_color(0xeb, 0x1a, 0x1d)
+        _make_color(120, 10, 12)
+        ; 1
+        ; B Button
+        _make_color(0xfe, 0xce, 0x15)
+        _make_color(136, 108, 0)
+        ; 2
+        ; X Button
+        _make_color(0x07, 0x49, 0xb4)
+        _make_color(0, 53, 144)
+        ; 3
+        ; Y Button
+        _make_color(0x00, 0x8d, 0x45)
+        _make_color(0, 70, 34)
+        ; 4
+        ; Select Buttons
+        .dw 0b0011110111101111
+        .dw 0b0001110011100111
+
+
+    update_palette:
+    """
+    Refresh menu/dialog palette entries from the configured colour at $16AA.
+    With ENABLE_BUTTON_DISPLAY, also pulls the action-button id and writes
+    the matching colour + shadow pair from `_button_colors` into $0CEF/$0CF1.
+    """
+        ldx.w 0x16AA
+        stx.w 0x0CDD
+        stx.w 0x0CE5
+    .if ENABLE_BUTTON_DISPLAY {
+        stx.w 0x0CED
+    ; black for the shadow but it could be a darker version of the color
+        ldx.w #0x0000
+        stx.w 0x0CEF
+        pha
+        jsr.w _get_action_button_id
+
+        cmp #0x04
+        bcc _proceed
+        ldx.w #5
+    _proceed:
+        asl
+        asl
+        tax
+
+        rep #0x20
+        lda.l _button_colors, x  ; color
+        sta.w 0x0CF1
+        lda.l _button_colors + 2, x  ; shadow color
+        sta.w 0x0CEF
+        sep #0x20
+        lda #0x00
+        xba
+        pla
+    }
     rtl
-    ;** routine principale
 
-vwfstart:
-"""
-Main entry for dialog VWF rendering: sets 8-bit A / 16-bit X-Y, enables HDMA, and falls into the parser loop
-that consumes the dialog text stream.
-"""
-    sep #0x20
-    rep #0x10
-
-    lda.b #0x01
-    sta.w 0x420D
-
-; 0x04-0x4F
-    var_base = 0x23
-    CNTR = var_base
-    CURRENT_C = var_base + 2
-    BITSLEFT = var_base + 4
-    font_addr = var_base + 6  ; 7 8
-    oldtilepos = var_base + 9
-    TILEPOS = var_base + 11
-    dialog_ptr = 0x20
-
-    no_wait_for_action = 0xcb
-
-    lda #0
-    jsr.w setup_font
-
-    php
-    sep #0x20
-.macro _clear_16_bit(var) {
-    stz.b var
-    stz.b var + 1
-}
-_clear_16_bit(CNTR)
-_clear_16_bit(CURRENT_C)
-_clear_16_bit(BITSLEFT)
-_clear_16_bit(oldtilepos)
-_clear_16_bit(TILEPOS)
-plp
+    window_palette:
+    """4 dialog/menu sub-palettes (4 colours each, BGR15)."""
+        ; Menu color palette I probably need to add one color
+        ; Palette 1, normal text
+        .db 0x00, 0x00
+        .db 0x00, 0x40
+        .db 0xCE, 0x39
+        .db 0xFF, 0x7F
+
+    ; Palette 2, greyed out text patched replaced the grey with black
+        .db 0x00, 0x00
+        .db 0x00, 0x40
+        .db 0x00, 0x00
+        .db 0xff, 0x7f
+
+    ; Palette  3, yellow
+        .db 0x00, 0x00
+        .db 0x00, 0x40
+        .db 0x80, 0x02
+        .db 0x7F, 0x03
+
+    ; Palette 4, red
+        .db 0x00, 0x00
+        .db 0x00, 0x40
+        .db 0xFF, 0x40
+        .db 0x7F, 0x2E
+
+    ;current_pos = 0x3d
+    ;current_text_pointer = 0x0772
+    ; 0x07 used for loop counter
+
+    ; ******************
+    ; ** Declarations **
+    ; ******************
+        vram_tile_set_pointer = 0x6800
+        vram_tile_map_pointer = 0x2C00
+        WRAM = field_vwf.tile_buffer
+
+        WRAMPTR = 0x2108
+
+    vwfinit:
+    """
+    One-shot dialog VWF initialisation: clear the WRAM scratch area,
+    DMA the VWF tile buffer + the original font tileset into VRAM,
+    and point BG3 at the new $6000 tile-set base.
+    """
+        jsr.w _clr  ; on efface un peu de Wram
+
+        jsr.w wait_for_vblank
+        dma_transfer_to_vram_call(WRAM, vram_tile_set_pointer, 0x0690, 0x1801)
+        jsr.w wait_for_vblank
+        dma_transfer_to_vram_call(WRAM, vram_tile_set_pointer + 0x348, 0x0690, 0x1801)
+
+    ; copy the old font tileset
+        jsr.w wait_for_vblank
+        dma_transfer_to_vram_call(0x0AF000, 0x6000, 0x800, 0x1801)
+        jsr.w wait_for_vblank
+        dma_transfer_to_vram_call(0x0AF000 + 0x800, 0x6000 + 0x400, 0x800, 0x1801)
+        jsr.w wait_for_vblank
+
+    ; Sets the BG3 vram pointer to 0x6000
+        lda 0x210C
+        and #0xF0
+        clc
+        adc #0x06
+        sta 0x210C
+
+        rtl
+        ;** routine principale
+
+    vwfstart:
+    """
+    Main entry for dialog VWF rendering: sets 8-bit A / 16-bit X-Y, enables HDMA, and falls into the parser loop
+    that consumes the dialog text stream.
+    """
+        sep #0x20
+        rep #0x10
+
+        lda.b #0x01
+        sta.w 0x420D
+
+    ; 0x04-0x4F
+        var_base = 0x23
+        CNTR = var_base
+        CURRENT_C = var_base + 2
+        BITSLEFT = var_base + 4
+        font_addr = var_base + 6  ; 7 8
+        oldtilepos = var_base + 9
+        TILEPOS = var_base + 11
+        dialog_ptr = 0x20
+
+        no_wait_for_action = 0xcb
+
+        lda #0
+        jsr.w setup_font
+
+        php
+        sep #0x20
+    .macro _clear_16_bit(var) {
+        stz.b var
+        stz.b var + 1
+    }
+    _clear_16_bit(CNTR)
+    _clear_16_bit(CURRENT_C)
+    _clear_16_bit(BITSLEFT)
+    _clear_16_bit(oldtilepos)
+    _clear_16_bit(TILEPOS)
+    plp
 
-lda.b #0x08
-sta.b BITSLEFT
+    lda.b #0x08
+    sta.b BITSLEFT
 
-jsr.l vwfinit
+    jsr.l vwfinit
 
-jsr.w load_letter
-bra _firstrun
+    jsr.w load_letter
+    bra _firstrun
 
-main:
-"""Per-character main loop."""
-    jsr.w load_letter_inc
+    main:
+    """Per-character main loop."""
+        jsr.w load_letter_inc
 
-_firstrun:
-    jmp.w parse
-    bra main
+    _firstrun:
+        jmp.w parse
+        bra main
 
-_fin:
-    ; this fixes a subtle issue where NPCs sprites positions and their collision table would be corrupted
-    xba
-    lda.b #0
-    xba
-    rtl
+    _fin:
+        ; this fixes a subtle issue where NPCs sprites positions and their collision table would be corrupted
+        xba
+        lda.b #0
+        xba
+        rtl
 
-;******************
-;** Parsing code **
-;******************
+    ;******************
+    ;** Parsing code **
+    ;******************
 
-parse:
-"""Dialog control-code dispatcher."""
+    parse:
+    """Dialog control-code dispatcher."""
 
-; Message Break
-    cmp #0x00
-    bne _nxt1
-    lda #0x01
-    sta 0xDE
-.if ENABLE_BUTTON_DISPLAY {
-    jsr.w _draw_button
-}
-jmp.w _fin
-
-_nxt1:
-    cmp #0x01
-    bne _nxt2
-    jmp.w newline
-
-_nxt2:
-    cmp #0x02
-    bne _nxt3
-    jmp.w space
-
-_nxt3:
-    ;Changement de Musique
-    cmp #0x03
-    bne _nxt4
-    jmp.w _musique
-
-_nxt4:
-    ; Nom des personages
-    cmp #0x04
-    bne _nxt5
-    jmp.w _display_character_name
-
-_nxt5:
-
-; wait
-    cmp #0x05
-    bne _nxt6
-    jmp.w _code05
-
-_nxt6:
-    ; Close window after dialog end
-    cmp #0x06
-    bne _nxt7
-    lda #0x02
-    sta 0xde
+    ; Message Break
+        cmp #0x00
+        bne _nxt1
+        lda #0x01
+        sta 0xDE
+    .if ENABLE_BUTTON_DISPLAY {
+        jsr.w _draw_button
+    }
     jmp.w _fin
-    ; Display item
-
-_nxt7:
-    cmp #0x07
-    bne _nxt8
-    jmp.w _code07
-
-; display gils count
-
-_nxt8:
-    cmp #0x08
-    bne _nxt_fb
-    jmp.w _code08
-
-_nxt_fb:
-    cmp #0xFC
-    bne _nxt_fc
-    jmp.w _suit3
-
-_nxt_fc:
-    cmp #0xFE
-    bne _nxt_fe
-    jsr.w load_letter_inc
-    jsr.w setup_font
-    jmp.w main
-
-_nxt_fe:
-
-    jsr.w _makeptr
-    jsr.w _shift_new
-    jsr.w _wdisplay
-
-    jmp.w main
-
-
-;***********
-;** Space **
-;***********
 
-space:
-"""Space control code ($02): advance tile cursor."""
-    jsr.w load_letter_inc
-    clc
-    adc.b TILEPOS
-    jmp.w main
+    _nxt1:
+        cmp #0x01
+        bne _nxt2
+        jmp.w newline
+
+    _nxt2:
+        cmp #0x02
+        bne _nxt3
+        jmp.w space
+
+    _nxt3:
+        ;Changement de Musique
+        cmp #0x03
+        bne _nxt4
+        jmp.w _musique
+
+    _nxt4:
+        ; Nom des personages
+        cmp #0x04
+        bne _nxt5
+        jmp.w _display_character_name
+
+    _nxt5:
+
+    ; wait
+        cmp #0x05
+        bne _nxt6
+        jmp.w _code05
+
+    _nxt6:
+        ; Close window after dialog end
+        cmp #0x06
+        bne _nxt7
+        lda #0x02
+        sta 0xde
+        jmp.w _fin
+        ; Display item
+
+    _nxt7:
+        cmp #0x07
+        bne _nxt8
+        jmp.w _code07
+
+    ; display gils count
+
+    _nxt8:
+        cmp #0x08
+        bne _nxt_fb
+        jmp.w _code08
+
+    _nxt_fb:
+        cmp #0xFC
+        bne _nxt_fc
+        jmp.w _suit3
+
+    _nxt_fc:
+        cmp #0xFE
+        bne _nxt_fe
+        jsr.w load_letter_inc
+        jsr.w setup_font
+        jmp.w main
+
+    _nxt_fe:
+
+        jsr.w _makeptr
+        jsr.w _shift_new
+        jsr.w _wdisplay
+
+        jmp.w main
+
+
+    ;***********
+    ;** Space **
+    ;***********
+
+    space:
+    """Space control code ($02): advance tile cursor."""
+        jsr.w load_letter_inc
+        clc
+        adc.b TILEPOS
+        jmp.w main
+
+    ;******************
+    ;Cout en gils
+    ;******************
+
+    _code08:
+        lda.w 0x08F8
+        sta.b 0x30
+        lda.w 0x08F9
+        sta.b 0x31
+        lda.w 0x08FA
+        sta.b 0x32
+        jsr.l 0x15C324
+
+        ldx.w #0x0000
+
+    _loop_b5c3:
+        lda.b 0x36, x
+        cmp #0x80
+        bne _loop_b5d2
+        inx
+        cpx.w #0x0005
+        beq _loop_b5d2
+        jmp.w _loop_b5c3
+
+    _loop_b5d2:
+        lda #0x00
+        xba
+        lda.b 0x36, x
+        ; Old code
+        ;00B5D4 STA 0x0774,Y
+        ;00B5D7 LDA #0xFF
+        ;00B5D9 STA 0x0834,Y
+        ;00B5DC INY
+        phx
+        phy
+
+        sta.b CURRENT_C  ; appel de la vwf
+
+        jsr.w _makeptr
+        jsr.w _shift_new
+        jsr.w _wdisplay
+
+        ply
+        plx
+
+        inx
+        cpx.w #0x0006
+        bne _loop_b5d2
+
+        jmp.w main
+
+    ;****************
+    ;** printfname **
+    ;****************
+
+    _display_character_name:
+    {
+        jsr.w load_letter_inc
+        xba
+        lda #0x00
+        xba
+
+        asl
+        pha
+        asl
+        clc
+        adc 1, s
+        tax
+        pla
+
+        ldy.w #0x0000
+    next:
+        lda #0x00
+        xba
+        lda 0x1500, x
+        sta.b CURRENT_C
+        cmp #0xFF
+        beq exit
+
+        phx
+        phy
+        jsr.w _makeptr
+        jsr.w _shift_new
+        jsr.w _wdisplay
+        ply
+        plx
+
+    suite:
+        inx
+
+        iny
+        cpy.w #0x0006
+        beq exit
+        jmp.w next
+    exit:
+        jmp.w main
+    }
+    ;********************
+    ;** Nouvelle ligne **
+    ;********************
 
-;******************
-;Cout en gils
-;******************
-
-_code08:
-    lda.w 0x08F8
-    sta.b 0x30
-    lda.w 0x08F9
-    sta.b 0x31
-    lda.w 0x08FA
-    sta.b 0x32
-    jsr.l 0x15C324
-
-    ldx.w #0x0000
-
-_loop_b5c3:
-    lda.b 0x36, x
-    cmp #0x80
-    bne _loop_b5d2
-    inx
-    cpx.w #0x0005
-    beq _loop_b5d2
-    jmp.w _loop_b5c3
-
-_loop_b5d2:
-    lda #0x00
-    xba
-    lda.b 0x36, x
-    ; Old code
-    ;00B5D4 STA 0x0774,Y
-    ;00B5D7 LDA #0xFF
-    ;00B5D9 STA 0x0834,Y
-    ;00B5DC INY
-    phx
-    phy
+    newline:
+    """Newline control code ($01): reset BITSLEFT and snap TILEPOS."""
+        rep #0x20
+        lda.w #8
+        sep #0x20
 
-    sta.b CURRENT_C  ; appel de la vwf
+        sta.b BITSLEFT
 
-    jsr.w _makeptr
-    jsr.w _shift_new
-    jsr.w _wdisplay
+        lda.b TILEPOS
+        clc
+        ;Second line
+        cmp #0x1A + 1
+        bcs suit
 
-    ply
-    plx
+        lda.b #0x1A
+        sta.b TILEPOS
+        bra end
 
-    inx
-    cpx.w #0x0006
-    bne _loop_b5d2
+    suit:
+    """Snap TILEPOS to line 3 ($34)."""
 
-    jmp.w main
+    ;Third Line
+        cmp #0x34 + 1
+        bcs _suit2
+        lda.b #0x34
+        sta.b TILEPOS
+        bra end
 
-;****************
-;** printfname **
-;****************
+    _suit2:
 
-_display_character_name:
-{
-    jsr.w load_letter_inc
-    xba
-    lda #0x00
-    xba
+    ;Forth Line
+        cmp #0x4E + 1
+        bcs _suit3
 
-    asl
-    pha
-    asl
-    clc
-    adc 1, s
-    tax
-    pla
+        lda.b #0x4E
+        sta.b TILEPOS
+        bra end
 
-    ldy.w #0x0000
-next:
-    lda #0x00
-    xba
-    lda 0x1500, x
-    sta.b CURRENT_C
-    cmp #0xFF
-    beq exit
-
-    phx
-    phy
-    jsr.w _makeptr
-    jsr.w _shift_new
-    jsr.w _wdisplay
-    ply
-    plx
-
-suite:
-    inx
-
-    iny
-    cpy.w #0x0006
-    beq exit
-    jmp.w next
-exit:
-    jmp.w main
-}
-;********************
-;** Nouvelle ligne **
-;********************
-
-newline:
-"""Newline control code ($01): reset BITSLEFT and snap TILEPOS."""
-    rep #0x20
-    lda.w #8
-    sep #0x20
+    _suit3:
+        ; that's where new ends up
 
+    .if ENABLE_BUTTON_DISPLAY {
+        jsr.w _draw_button
+    }
+    lda.b #0x08
     sta.b BITSLEFT
 
-    lda.b TILEPOS
-    clc
-    ;Second line
-    cmp #0x1A + 1
-    bcs suit
-
-    lda.b #0x1A
-    sta.b TILEPOS
-    bra end
-
-suit:
-"""Snap TILEPOS to line 3 ($34)."""
-
-;Third Line
-    cmp #0x34 + 1
-    bcs _suit2
-    lda.b #0x34
-    sta.b TILEPOS
-    bra end
-
-_suit2:
-
-;Forth Line
-    cmp #0x4E + 1
-    bcs _suit3
-
-    lda.b #0x4E
-    sta.b TILEPOS
-    bra end
-
-_suit3:
-    ; that's where new ends up
+    stz.b CURRENT_C
+    stz.b TILEPOS
+    jsr.w _clr
+    jsr.w _waitpad
+    jsr.w _wdisplay
 
-.if ENABLE_BUTTON_DISPLAY {
-    jsr.w _draw_button
-}
-lda.b #0x08
-sta.b BITSLEFT
+    end:
+    """Common newline tail."""
 
-stz.b CURRENT_C
-stz.b TILEPOS
-jsr.w _clr
-jsr.w _waitpad
-jsr.w _wdisplay
+        jmp.w main
 
-end:
-"""Common newline tail."""
+    _draw_button:
+    {
+        jsr.w _get_action_button_id
+        cmp #4  ; select
+        bne _normal_button
+        lda #0xa5
+        bra _store
+    _normal_button:
+        lda #0xa4
+    _store:
+        sta.b CURRENT_C
+        jsr.w _draw_ending_symbol
+
+        rts
+    }
 
+    _draw_ending_symbol:
+        pha
+
+        lda #0x08
+        sta.b BITSLEFT
+        lda #0x4E + 24
+        sta.b oldtilepos
+
+        lda #0x4E + 25
+        sta.b TILEPOS
+
+
+        ldx.w #0xa2 * 17
+        ldy.w #0xcc0
+        jsr.w _makeptr
+        jsr.w _shift_new
+        jsr.w _wdisplay
+
+        pla
+        rts
+
+
+    ;*************
+    ;** Musique **
+    ;*************
+
+    _musique:
+        jsr.w load_letter_inc
+        sta 0x1E01
+        lda.b #0x01
+        sta 0x1E00
+
+        jsr.l 0x048004
+        jmp.w main
+
+    _code05:
+        jsr.w load_letter_inc
+        xba
+        lda #0x00
+        xba
+        asl
+        asl
+        asl
+        tax
+        stx 0x08F4
+        ldx 0x0000
+        stx 0x08F6
+    {
+        ldx 0x08f4
+        beq skip
+    loop:
+        cpx 0x08f6
+
+        bne loop
+    skip:
+        ldx.w #0x0000
+        stx 0x08f4
+    }
     jmp.w main
 
-_draw_button:
-{
-    jsr.w _get_action_button_id
-    cmp #4  ; select
-    bne _normal_button
-    lda #0xa5
-    bra _store
-_normal_button:
-    lda #0xa4
-_store:
-    sta.b CURRENT_C
-    jsr.w _draw_ending_symbol
 
-    rts
-}
+    .macro _vwf_putchar() {
+        phx
+        phy
+        sta.b CURRENT_C
+        jsr.w _makeptr
+        jsr.w _shift_new
+        jsr.w _wdisplay
+        ply
+        plx
+    }
 
-_draw_ending_symbol:
-    pha
+    ; display item or magic
 
-    lda #0x08
-    sta.b BITSLEFT
-    lda #0x4E + 24
-    sta.b oldtilepos
+    _code07:
+    {
+        lda 0x08FB
+
+        rep #0x20
+        and.w #0x00FF
+        pha
+        clc
+        adc 0x01, s  ; x2
+        adc 0x01, s  ; x3
+        asl
+        asl  ; x12
+        tax
+        pla
+        sep #0x20
+
+    ; skip first char (usually a space or a symbol.)
+        inx
+        lda #0x0b  ; 11 characters
+
+    loop:
+        pha
+        lda.b #0x00
+        xba
+        lda.l assets_items_dat, x
+        cmp #0xFF
+        beq cleanup
+        _vwf_putchar()
+
+        inx
+
+        pla
+        dec
+        bne loop
+        bra end
+    cleanup:
+        pla
+    end:
+        lda #0x00
+        jmp.w main
+    }
 
-    lda #0x4E + 25
-    sta.b TILEPOS
+    ;*******************
+    ;** Shift Routine **
+    ;*******************
+
+    _shift_new:
+        rep #0x20
+        lda.w #0x0010
+        sta.b CNTR
+        sep #0x20
+
+    _boucle2:
+        rep #0x20
+        lda.w #0x0000
+        sep #0x20
+        phx
+        lda.b BITSLEFT
+
+        cmp #0x08
+        bne _shift
+        plx
+        ; LDA.L assets_font_dat,X
+        phy
+        txy
+        lda.b [font_addr], y
+        tyx
+        ply
+        inx
+        xba
+        bra _store
+
+    _shift:
+    .if 1 {
+        plx
+        phy
+        txy
+        lda.b [font_addr], y
+        tyx
+        ply
+
+        xba
+        lda #0x00
+        xba
+
+    ; make jump_table_pointer
+        phx
+        pha
+        lda.b BITSLEFT
+        asl
+        tax
+        pla
+
+        rep #0x20
+        jmp.w (_mul_table, x)
+    _mul_table:
+        .dw _mul_0
+        .dw _mul_1
+        .dw _mul_2
+        .dw _mul_3
+        .dw _mul_4
+        .dw _mul_5
+        .dw _mul_6
+        .dw _mul_7
+        .dw _mul_8
+
+    _mul_8:
+    _mul_7:
+        asl  ; 1
+    _mul_6:
+        asl  ; 2
+    _mul_5:
+        asl  ; 3
+    _mul_4:
+        asl  ; 4
+    _mul_3:
+        asl  ; 5
+    _mul_2:
+        asl  ; 6
+    _mul_1:
+        asl  ; 7
+    _mul_0:
+        plx
+        sep #0x20
+        inx
+    } else {
+        tax  ; using math multiplication
+        lda.l vwf_shift_table, x
+        sta.l 0x004202  ; MULTPILIER
 
 
-    ldx.w #0xa2 * 17
-    ldy.w #0xcc0
-    jsr.w _makeptr
-    jsr.w _shift_new
-    jsr.w _wdisplay
+        plx
 
-    pla
-    rts
+        phy
+        txy
+        lda.b [font_addr], y
+        tyx
+        ply
+        inx
 
+        sta.l 0x004203  ; MULTIPLICAND
 
-;*************
-;** Musique **
-;*************
+        rep #0x20
+        nop
+        nop
+        nop
+        nop
+        lda.l 0x004216  ; the result is stored in 0x4216-0x4217
+        sep #0x20
+    }
 
-_musique:
-    jsr.w load_letter_inc
-    sta 0x1E01
-    lda.b #0x01
-    sta 0x1E00
+    _store:
 
-    jsr.l 0x048004
-    jmp.w main
+    ; ff
+    ; notre bidule
+    ; 0b11111111 A
+    ; 0b00000001 B
+    ; roll B and B xor A ?
+    ; 0b11 = white
+    ; 0b10 = black
+    ; 0b01 = window background
+    ; 0b00 = transparent
+    TEXT_SHADOW := 1
 
-_code05:
-    jsr.w load_letter_inc
     xba
-    lda #0x00
-    xba
-    asl
-    asl
-    asl
-    tax
-    stx 0x08F4
-    ldx 0x0000
-    stx 0x08F6
-{
-    ldx 0x08f4
-    beq skip
-loop:
-    cpx 0x08f6
-
-    bne loop
-skip:
-    ldx.w #0x0000
-    stx 0x08f4
-}
-jmp.w main
-
-
-.macro _vwf_putchar() {
     phx
-    phy
-    sta.b CURRENT_C
-    jsr.w _makeptr
-    jsr.w _shift_new
-    jsr.w _wdisplay
-    ply
-    plx
-}
-
-; display item or magic
+    tyx
 
-_code07:
-{
-    lda 0x08FB
 
-    rep #0x20
-    and.w #0x00FF
     pha
-    clc
-    adc 0x01, s  ; x2
-    adc 0x01, s  ; x3
-    asl
-    asl  ; x12
-    tax
+    ora.l WRAM, x
+    sta.l WRAM, x
     pla
-    sep #0x20
 
-; skip first char (usually a space or a symbol.)
-    inx
-    lda #0x0b  ; 11 characters
-
-loop:
     pha
-    lda.b #0x00
-    xba
-    lda.l assets_items_dat, x
-    cmp #0xFF
-    beq cleanup
-    _vwf_putchar()
+    ora.l WRAM + 1, x
+    sta.l WRAM + 1, x
+    pla
 
-    inx
+    .if TEXT_SHADOW {
+        pha
+        eor.l WRAM + 3, x
+        sta.l WRAM + 3, x
+        pla
+        pha
+        eor.l WRAM + 2, x
+        sta.l WRAM + 2, x
+        pla
+    }
+    xba
 
+    pha
+    ora.l WRAM + 0x20, x
+    sta.l WRAM + 0x20, x
     pla
-    dec
-    bne loop
-    bra end
-cleanup:
-    pla
-end:
-    lda #0x00
-    jmp.w main
-}
 
-;*******************
-;** Shift Routine **
-;*******************
-
-_shift_new:
-    rep #0x20
-    lda.w #0x0010
-    sta.b CNTR
-    sep #0x20
+    pha
+    ora.l WRAM + 0x20 + 1, x
+    sta.l WRAM + 0x20 + 1, x
+    pla
 
-_boucle2:
-    rep #0x20
-    lda.w #0x0000
-    sep #0x20
-    phx
-    lda.b BITSLEFT
+    .if TEXT_SHADOW {
+        pha
+        eor.l WRAM + 3 + 0x20, x
+        sta.l WRAM + 3 + 0x20, x
+        pla
 
-    cmp #0x08
-    bne _shift
-    plx
-    ; LDA.L assets_font_dat,X
-    phy
+        pha
+        eor.l WRAM + 2 + 0x20, x
+        sta.l WRAM + 2 + 0x20, x
+        pla
+    }
     txy
-    lda.b [font_addr], y
-    tyx
-    ply
-    inx
-    xba
-    bra _store
-
-_shift:
-.if 1 {
     plx
-    phy
-    txy
-    lda.b [font_addr], y
-    tyx
-    ply
 
-    xba
-    lda #0x00
-    xba
-
-; make jump_table_pointer
-    phx
-    pha
-    lda.b BITSLEFT
-    asl
-    tax
-    pla
+    iny
+    iny
 
-    rep #0x20
-    jmp.w (_mul_table, x)
-_mul_table:
-    .dw _mul_0
-    .dw _mul_1
-    .dw _mul_2
-    .dw _mul_3
-    .dw _mul_4
-    .dw _mul_5
-    .dw _mul_6
-    .dw _mul_7
-    .dw _mul_8
-
-_mul_8:
-_mul_7:
-    asl  ; 1
-_mul_6:
-    asl  ; 2
-_mul_5:
-    asl  ; 3
-_mul_4:
-    asl  ; 4
-_mul_3:
-    asl  ; 5
-_mul_2:
-    asl  ; 6
-_mul_1:
-    asl  ; 7
-_mul_0:
-    plx
-    sep #0x20
-    inx
-} else {
-    tax  ; using math multiplication
-    lda.l vwf_shift_table, x
-    sta.l 0x004202  ; MULTPILIER
+    dec.b CNTR
+    beq _exit
+    jmp.w _boucle2
 
+    _exit:
 
-    plx
+        rep #0x20
+    .if ENABLE_KERNING {
+    ; Routine reads the pair from CURRENT_C (DP), no need to load A.
+        jsr.w dialog_get_kerning_adjustment_binary_search
+        pha
+    } else {
+        lda.w #0x0000
+        pha
+    }
 
-    phy
     txy
     lda.b [font_addr], y
     tyx
-    ply
-    inx
 
-    sta.l 0x004203  ; MULTIPLICAND
 
-    rep #0x20
-    nop
-    nop
-    nop
-    nop
-    lda.l 0x004216  ; the result is stored in 0x4216-0x4217
     sep #0x20
-}
-
-_store:
-
-; ff
-; notre bidule
-; 0b11111111 A
-; 0b00000001 B
-; roll B and B xor A ?
-; 0b11 = white
-; 0b10 = black
-; 0b01 = window background
-; 0b00 = transparent
-TEXT_SHADOW := 1
-
-xba
-phx
-tyx
-
-
-pha
-ora.l WRAM, x
-sta.l WRAM, x
-pla
-
-pha
-ora.l WRAM + 1, x
-sta.l WRAM + 1, x
-pla
-
-.if TEXT_SHADOW {
-    pha
-    eor.l WRAM + 3, x
-    sta.l WRAM + 3, x
-    pla
-    pha
-    eor.l WRAM + 2, x
-    sta.l WRAM + 2, x
-    pla
-}
-xba
-
-pha
-ora.l WRAM + 0x20, x
-sta.l WRAM + 0x20, x
-pla
-
-pha
-ora.l WRAM + 0x20 + 1, x
-sta.l WRAM + 0x20 + 1, x
-pla
-
-.if TEXT_SHADOW {
     pha
-    eor.l WRAM + 3 + 0x20, x
-    sta.l WRAM + 3 + 0x20, x
-    pla
-
-    pha
-    eor.l WRAM + 2 + 0x20, x
-    sta.l WRAM + 2 + 0x20, x
-    pla
-}
-txy
-plx
-
-iny
-iny
-
-dec.b CNTR
-beq _exit
-jmp.w _boucle2
-
-_exit:
-
-    rep #0x20
-.if ENABLE_KERNING {
-; Routine reads the pair from CURRENT_C (DP), no need to load A.
-    jsr.w dialog_get_kerning_adjustment_binary_search
-    pha
-} else {
-    lda.w #0x0000
-    pha
-}
-
-txy
-lda.b [font_addr], y
-tyx
 
+    lda.b BITSLEFT
 
-sep #0x20
-pha
-
-lda.b BITSLEFT
-
-clc
-sbc.b 0x01, s
-
-clc
-adc.b 0x02, s
-
-loopdec:
-"""Tail of the per-glyph blit."""
-    cmp #0x00
-    bmi coupe
-    beq coupe
-
-    sta.b BITSLEFT
-    pla
-    pla
-    pla
-    rts
-
-coupe:
-"""Glyph crossed a tile boundary: bump TILEPOS."""
     clc
-    adc #0x08
-    inc.b TILEPOS
-    bra loopdec
-
-
-vwf_shift_table:
-"""8-bit shift mask table indexed by BITSLEFT."""
-    .db 0b00000000  ; dummy entrie =0
-    .db 0b00000010  ; 1
-    .db 0b00000100  ; 2
-    .db 0b00001000  ; 3
-    .db 0b00010000  ; 4
-    .db 0b00100000  ; 5
-    .db 0b01000000  ; 6
-    .db 0b10000000  ; 7
-    .db 0b10000000  ; 8
-
-; Long-form (RTL) wrapper for cross-bank callers and Python tests.
-
-setup_font_ext:
-"""RTL wrapper around setup_font for cross-bank callers."""
-    jsr.w setup_font
-    rtl
+    sbc.b 0x01, s
 
-setup_font:
-"""Load the 24-bit address of font A from font_table."""
-    ; A: font index
-{
-    phx
-    pha
-    asl
     clc
-    adc 0x01, s
-    xba
-    lda #0
-    xba
-    tax
-    pla
-
-    rep #0x20
-    lda.l font_table, x
-    sta.b font_addr
-    sep #0x20
-
-    lda.l font_table + 2, x
-    sta.b font_addr + 2
-    plx
-    rts
-}
-
-;************************
-;** build font pointer **
-;************************
-
-_makeptr:
-    pha
-
-    ldx.w #0x0000
-    ldy.w #0x0000
-    lda.b CURRENT_C
-    xba
-    lda #0x00
-    xba
-    rep #0x20
-    pha
-    asl
-    asl
-    asl
-    asl
-    adc.b 0x01, s
-    tax
-    pla
-    lda.w #0x0000
-    sep #0x20
-
-    lda.b TILEPOS
-    sta.b oldtilepos
-    rep #0x20
-    asl
-    asl
-    asl
-    asl
-    asl
-    sta.b oldtilepos
-    tay
-    sep #0x20
-    pla
-    rts
-
-
-;===================================
-;Clear Wram
-;
-;===================================
-
-_clr:
-{
-    phx
-    ldx.w #0x0000
-solid_bg_loop:
-    lda.b #0xff
-    sta.l WRAM, x
-    inx
-
-    lda.b #0x00
-    sta.l WRAM, x
-    inx
-    cpx.w #0x0D10
-
-    bne solid_bg_loop
-
-
-transparent_bg_loop:
-    lda.b #0x00
-    sta.l WRAM, x
-    inx
-    cpx.w #0x0D20
-    bne transparent_bg_loop
-
-    plx
-    rts
-}
-
-_wait_key_up:
-    lda.l 0x000602
-    bne _wait_key_up
-    lda.l 0x000603
-    bne _wait_key_up
-    rts
-
-_wait_key_down:
-{
-    ACTION_BUTTON := 0x80
-    lda.l 0x000602
-    bit #ACTION_BUTTON  ; Action button
-    bne exit
-    wai
-    bra _wait_key_down
-exit:
-    rts
-}
+    adc.b 0x02, s
+
+    loopdec:
+    """Tail of the per-glyph blit."""
+        cmp #0x00
+        bmi coupe
+        beq coupe
+
+        sta.b BITSLEFT
+        pla
+        pla
+        pla
+        rts
+
+    coupe:
+    """Glyph crossed a tile boundary: bump TILEPOS."""
+        clc
+        adc #0x08
+        inc.b TILEPOS
+        bra loopdec
+
+
+    vwf_shift_table:
+    """8-bit shift mask table indexed by BITSLEFT."""
+        .db 0b00000000  ; dummy entrie =0
+        .db 0b00000010  ; 1
+        .db 0b00000100  ; 2
+        .db 0b00001000  ; 3
+        .db 0b00010000  ; 4
+        .db 0b00100000  ; 5
+        .db 0b01000000  ; 6
+        .db 0b10000000  ; 7
+        .db 0b10000000  ; 8
+
+    ; Long-form (RTL) wrapper for cross-bank callers and Python tests.
+
+    setup_font_ext:
+    """RTL wrapper around setup_font for cross-bank callers."""
+        jsr.w setup_font
+        rtl
+
+    setup_font:
+    """Load the 24-bit address of font A from font_table."""
+        ; A: font index
+    {
+        phx
+        pha
+        asl
+        clc
+        adc 0x01, s
+        xba
+        lda #0
+        xba
+        tax
+        pla
+
+        rep #0x20
+        lda.l font_table, x
+        sta.b font_addr
+        sep #0x20
+
+        lda.l font_table + 2, x
+        sta.b font_addr + 2
+        plx
+        rts
+    }
 
-;Wait for joypad 1
+    ;************************
+    ;** build font pointer **
+    ;************************
+
+    _makeptr:
+        pha
+
+        ldx.w #0x0000
+        ldy.w #0x0000
+        lda.b CURRENT_C
+        xba
+        lda #0x00
+        xba
+        rep #0x20
+        pha
+        asl
+        asl
+        asl
+        asl
+        adc.b 0x01, s
+        tax
+        pla
+        lda.w #0x0000
+        sep #0x20
+
+        lda.b TILEPOS
+        sta.b oldtilepos
+        rep #0x20
+        asl
+        asl
+        asl
+        asl
+        asl
+        sta.b oldtilepos
+        tay
+        sep #0x20
+        pla
+        rts
+
+
+    ;===================================
+    ;Clear Wram
+    ;
+    ;===================================
+
+    _clr:
+    {
+        phx
+        ldx.w #0x0000
+    solid_bg_loop:
+        lda.b #0xff
+        sta.l WRAM, x
+        inx
+
+        lda.b #0x00
+        sta.l WRAM, x
+        inx
+        cpx.w #0x0D10
+
+        bne solid_bg_loop
+
+
+    transparent_bg_loop:
+        lda.b #0x00
+        sta.l WRAM, x
+        inx
+        cpx.w #0x0D20
+        bne transparent_bg_loop
+
+        plx
+        rts
+    }
 
-_waitpad:
-{
-    pha
-    lda.b no_wait_for_action
-    bne nowaitpad
+    _wait_key_up:
+        lda.l 0x000602
+        bne _wait_key_up
+        lda.l 0x000603
+        bne _wait_key_up
+        rts
 
-    .if ENABLE_DIALOG_SKIP {
-    jsr.w _wait_key_up
+    _wait_key_down:
+    {
+        ACTION_BUTTON := 0x80
+        lda.l 0x000602
+        bit #ACTION_BUTTON  ; Action button
+        bne exit
+        wai
+        bra _wait_key_down
+    exit:
+        rts
     }
-    jsr.w _wait_key_down
-    bra end
 
+    ;Wait for joypad 1
 
-nowaitpad:
-    lda.b #0x30
+    _waitpad:
     {
-loop:
-    jsr.w wait_for_vblank
-    dec
-    bne loop
+        pha
+        lda.b no_wait_for_action
+        bne nowaitpad
+
+        .if ENABLE_DIALOG_SKIP {
+        jsr.w _wait_key_up
+        }
+        jsr.w _wait_key_down
+        bra end
+
+
+    nowaitpad:
+        lda.b #0x30
+        {
+    loop:
+        jsr.w wait_for_vblank
+        dec
+        bne loop
+        }
+    end:
+        pla
+        jsr.w _clr
+        jsr.w wait_for_vblank
+        dma_transfer_to_vram_call(WRAM, vram_tile_set_pointer, 0x0690, 0x1801)
+        jsr.w wait_for_vblank
+        dma_transfer_to_vram_call(WRAM + 0x348, vram_tile_set_pointer + 0x348, 0x0690, 0x1801)
+        rts
     }
-end:
-    pla
-    jsr.w _clr
-    jsr.w wait_for_vblank
-    dma_transfer_to_vram_call(WRAM, vram_tile_set_pointer, 0x0690, 0x1801)
-    jsr.w wait_for_vblank
-    dma_transfer_to_vram_call(WRAM + 0x348, vram_tile_set_pointer + 0x348, 0x0690, 0x1801)
-    rts
-}
 
-_wdisplay:
-    ; wait for vblank to transfer
-    jsr.w wait_for_vblank
+    _wdisplay:
+        ; wait for vblank to transfer
+        jsr.w wait_for_vblank
 
-    sep #0x20
+        sep #0x20
 
 
-;macro expansion
+    ;macro expansion
 
-    php
-    pha
-    phx
+        php
+        pha
+        phx
 
-    lda.b #0x80
-    sta.w 0x2115
+        lda.b #0x80
+        sta.w 0x2115
 
-    rep #0x20
-    pha
+        rep #0x20
+        pha
 
-    lda.b oldtilepos
-    lsr  ; addresse vram /2
-    clc
-    adc.w #vram_tile_set_pointer
-    sta.w 0x2116
+        lda.b oldtilepos
+        lsr  ; addresse vram /2
+        clc
+        adc.w #vram_tile_set_pointer
+        sta.w 0x2116
 
 
-    lda.b oldtilepos
-    clc
-    adc.w #WRAM & 0xFFFF
-    sta.w 0x4372
+        lda.b oldtilepos
+        clc
+        adc.w #WRAM & 0xFFFF
+        sta.w 0x4372
 
-    pla
-    sep #0x20
+        pla
+        sep #0x20
 
-    channel = 7
+        channel = 7
 
-    ldx.w #0x1801
-    stx.w 0x4370
-    lda.b #0xFF & ( WRAM >> 16 )
-    sta.w 0x4374
+        ldx.w #0x1801
+        stx.w 0x4370
+        lda.b #0xFF & ( WRAM >> 16 )
+        sta.w 0x4374
 
-    ldx.w #0x0040
-    stx.w 0x4375
-    lda.b #0x01 << 7
-    sta 0x420B
+        ldx.w #0x0040
+        stx.w 0x4375
+        lda.b #0x01 << 7
+        sta 0x420B
 
-    nop
-    nop
+        nop
+        nop
 
-    plx
-    pla
-    plp
+        plx
+        pla
+        plp
 
 
-    lda.b no_wait_for_action
-    bne _nowindow
+        lda.b no_wait_for_action
+        bne _nowindow
 
-    dma_transfer_to_vram_call(_winmap, vram_tile_map_pointer, _endwinmap - _winmap, 0x1801)
-    bra window
+        dma_transfer_to_vram_call(_winmap, vram_tile_map_pointer, _endwinmap - _winmap, 0x1801)
+        bra window
 
-_nowindow:
-    dma_transfer_to_vram_call(_intromap, vram_tile_map_pointer, _endintromap - _intromap, 0x1801)
+    _nowindow:
+        dma_transfer_to_vram_call(_intromap, vram_tile_map_pointer, _endintromap - _intromap, 0x1801)
 
-window:
-"""Post-DMA window-display tail."""
-{
-    lda.b CURRENT_C
-    cmp #0xFF
-    beq no_char_wait
-    wai  ; wait for interrupts
-    lda.b no_wait_for_action  ; add extra wait when in "story telling".
-    beq no_char_wait
-    jsr.w wait_for_vblank
-    jsr.w wait_for_vblank
-no_char_wait:
-}
-rts
+    window:
+    """Post-DMA window-display tail."""
+    {
+        lda.b CURRENT_C
+        cmp #0xFF
+        beq no_char_wait
+        wai  ; wait for interrupts
+        lda.b no_wait_for_action  ; add extra wait when in "story telling".
+        beq no_char_wait
+        jsr.w wait_for_vblank
+        jsr.w wait_for_vblank
+    no_char_wait:
+    }
+    rts
 
 
-_winmap:
-.if TEXT_SHADOW {
-    .if ENABLE_BUTTON_DISPLAY {
-    last_character_palette := 2
+    _winmap:
+    .if TEXT_SHADOW {
+        .if ENABLE_BUTTON_DISPLAY {
+        last_character_palette := 2
+        } else {
+        last_character_palette := 1
+        }
+        .dw 0x2000, 0x2000, 0x2019, 0x2500, 0x2502, 0x2504, 0x2506, 0x2508, 0x250A, 0x250C, 0x250E, 0x2510, 0x2512, 0x2514, 0x2516, 0x2518, 0x251A, 0x251C, 0x251E, 0x2520, 0x2522, 0x2524, 0x2526, 0x2528, 0x252A, 0x252C, 0x252E, 0x2530, 0x2532, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2501, 0x2503, 0x2505, 0x2507, 0x2509, 0x250B, 0x250D, 0x250F, 0x2511, 0x2513, 0x2515, 0x2517, 0x2519, 0x251B, 0x251D, 0x251F, 0x2521, 0x2523, 0x2525, 0x2527, 0x2529, 0x252B, 0x252D, 0x252F, 0x2531, 0x2533, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2534, 0x2536, 0x2538, 0x253A, 0x253C, 0x253E, 0x2540, 0x2542, 0x2544, 0x2546, 0x2548, 0x254A, 0x254C, 0x254E, 0x2550, 0x2552, 0x2554, 0x2556, 0x2558, 0x255A, 0x255C, 0x255E, 0x2560, 0x2562, 0x2564, 0x2566, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2535, 0x2537, 0x2539, 0x253B, 0x253D, 0x253F, 0x2541, 0x2543, 0x2545, 0x2547, 0x2549, 0x254B, 0x254D, 0x254F, 0x2551, 0x2553, 0x2555, 0x2557, 0x2559, 0x255B, 0x255D, 0x255F, 0x2561, 0x2563, 0x2565, 0x2567, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2568, 0x256A, 0x256C, 0x256E, 0x2570, 0x2572, 0x2574, 0x2576, 0x2578, 0x257A, 0x257C, 0x257E, 0x2580, 0x2582, 0x2584, 0x2586, 0x2588, 0x258A, 0x258C, 0x258E, 0x2590, 0x2592, 0x2594, 0x2596, 0x2598, 0x259A, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2569, 0x256B, 0x256D, 0x256F, 0x2571, 0x2573, 0x2575, 0x2577, 0x2579, 0x257B, 0x257D, 0x257F, 0x2581, 0x2583, 0x2585, 0x2587, 0x2589, 0x258B, 0x258D, 0x258F, 0x2591, 0x2593, 0x2595, 0x2597, 0x2599, 0x259B, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x259C, 0x259E, 0x25A0, 0x25A2, 0x25A4, 0x25A6, 0x25A8, 0x25AA, 0x25AC, 0x25AE, 0x25B0, 0x25B2, 0x25B4, 0x25B6, 0x25B8, 0x25BA, 0x25BC, 0x25BE, 0x25C0, 0x25C2, 0x25C4, 0x25C6, 0x25C8, 0x25CA, 0x25CC, ( 0x21CE + ( last_character_palette << 10 ) ), 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x259D, 0x259F, 0x25A1, 0x25A3, 0x25A5, 0x25A7, 0x25A9, 0x25AB, 0x25AD, 0x25AF, 0x25B1, 0x25B3, 0x25B5, 0x25B7, 0x25B9, 0x25BB, 0x25BD, 0x25BF, 0x25C1, 0x25C3, 0x25C5, 0x25C7, 0x25C9, 0x25CB, 0x25CD, ( 0x21CF + ( last_character_palette << 10 ) ), 0x201A, 0x2000, 0x2000  ; noqa: E501
     } else {
-    last_character_palette := 1
+        .dw 0x2000, 0x2000, 0x2019, 0x2100, 0x2102, 0x2104, 0x2106, 0x2108, 0x210A, 0x210C, 0x210E, 0x2110, 0x2112, 0x2114, 0x2116, 0x2118, 0x211A, 0x211C, 0x211E, 0x2120, 0x2122, 0x2124, 0x2126, 0x2128, 0x212A, 0x212C, 0x212E, 0x2130, 0x2132, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2101, 0x2103, 0x2105, 0x2107, 0x2109, 0x210B, 0x210D, 0x210F, 0x2111, 0x2113, 0x2115, 0x2117, 0x2119, 0x211B, 0x211D, 0x211F, 0x2121, 0x2123, 0x2125, 0x2127, 0x2129, 0x212B, 0x212D, 0x212F, 0x2131, 0x2133, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2134, 0x2136, 0x2138, 0x213A, 0x213C, 0x213E, 0x2140, 0x2142, 0x2144, 0x2146, 0x2148, 0x214A, 0x214C, 0x214E, 0x2150, 0x2152, 0x2154, 0x2156, 0x2158, 0x215A, 0x215C, 0x215E, 0x2160, 0x2162, 0x2164, 0x2166, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2135, 0x2137, 0x2139, 0x213B, 0x213D, 0x213F, 0x2141, 0x2143, 0x2145, 0x2147, 0x2149, 0x214B, 0x214D, 0x214F, 0x2151, 0x2153, 0x2155, 0x2157, 0x2159, 0x215B, 0x215D, 0x215F, 0x2161, 0x2163, 0x2165, 0x2167, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2168, 0x216A, 0x216C, 0x216E, 0x2170, 0x2172, 0x2174, 0x2176, 0x2178, 0x217A, 0x217C, 0x217E, 0x2180, 0x2182, 0x2184, 0x2186, 0x2188, 0x218A, 0x218C, 0x218E, 0x2190, 0x2192, 0x2194, 0x2196, 0x2198, 0x219A, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x2169, 0x216B, 0x216D, 0x216F, 0x2171, 0x2173, 0x2175, 0x2177, 0x2179, 0x217B, 0x217D, 0x217F, 0x2181, 0x2183, 0x2185, 0x2187, 0x2189, 0x218B, 0x218D, 0x218F, 0x2191, 0x2193, 0x2195, 0x2197, 0x2199, 0x219B, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x219C, 0x219E, 0x21A0, 0x21A2, 0x21A4, 0x21A6, 0x21A8, 0x21AA, 0x21AC, 0x21AE, 0x21B0, 0x21B2, 0x21B4, 0x21B6, 0x21B8, 0x21BA, 0x21BC, 0x21BE, 0x21C0, 0x21C2, 0x21C4, 0x21C6, 0x21C8, 0x21CA, 0x21CC, 0x21CE, 0x201A, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2019, 0x219D, 0x219F, 0x21A1, 0x21A3, 0x21A5, 0x21A7, 0x21A9, 0x21AB, 0x21AD, 0x21AF, 0x21B1, 0x21B3, 0x21B5, 0x21B7, 0x21B9, 0x21BB, 0x21BD, 0x21BF, 0x21C1, 0x21C3, 0x21C5, 0x21C7, 0x21C9, 0x21CB, 0x21CD, 0x21CF, 0x201A, 0x2000, 0x2000  ; noqa: E501
     }
-    .dw 0x2000, 0x2000, 0x2019, 0x2500, 0x2502, 0x2504, 0x2506, 0x2508, 0x250A, 0x250C, 0x250E, 0x2510, 0x2512, 0x2514, 0x2516, 0x2518, 0x251A, 0x251C, 0x251E, 0x2520, 0x2522, 0x2524, 0x2526, 0x2528, 0x252A, 0x252C, 0x252E, 0x2530, 0x2532, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2501, 0x2503, 0x2505, 0x2507, 0x2509, 0x250B, 0x250D, 0x250F, 0x2511, 0x2513, 0x2515, 0x2517, 0x2519, 0x251B, 0x251D, 0x251F, 0x2521, 0x2523, 0x2525, 0x2527, 0x2529, 0x252B, 0x252D, 0x252F, 0x2531, 0x2533, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2534, 0x2536, 0x2538, 0x253A, 0x253C, 0x253E, 0x2540, 0x2542, 0x2544, 0x2546, 0x2548, 0x254A, 0x254C, 0x254E, 0x2550, 0x2552, 0x2554, 0x2556, 0x2558, 0x255A, 0x255C, 0x255E, 0x2560, 0x2562, 0x2564, 0x2566, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2535, 0x2537, 0x2539, 0x253B, 0x253D, 0x253F, 0x2541, 0x2543, 0x2545, 0x2547, 0x2549, 0x254B, 0x254D, 0x254F, 0x2551, 0x2553, 0x2555, 0x2557, 0x2559, 0x255B, 0x255D, 0x255F, 0x2561, 0x2563, 0x2565, 0x2567, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2568, 0x256A, 0x256C, 0x256E, 0x2570, 0x2572, 0x2574, 0x2576, 0x2578, 0x257A, 0x257C, 0x257E, 0x2580, 0x2582, 0x2584, 0x2586, 0x2588, 0x258A, 0x258C, 0x258E, 0x2590, 0x2592, 0x2594, 0x2596, 0x2598, 0x259A, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2569, 0x256B, 0x256D, 0x256F, 0x2571, 0x2573, 0x2575, 0x2577, 0x2579, 0x257B, 0x257D, 0x257F, 0x2581, 0x2583, 0x2585, 0x2587, 0x2589, 0x258B, 0x258D, 0x258F, 0x2591, 0x2593, 0x2595, 0x2597, 0x2599, 0x259B, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x259C, 0x259E, 0x25A0, 0x25A2, 0x25A4, 0x25A6, 0x25A8, 0x25AA, 0x25AC, 0x25AE, 0x25B0, 0x25B2, 0x25B4, 0x25B6, 0x25B8, 0x25BA, 0x25BC, 0x25BE, 0x25C0, 0x25C2, 0x25C4, 0x25C6, 0x25C8, 0x25CA, 0x25CC, ( 0x21CE + ( last_character_palette << 10 ) ), 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x259D, 0x259F, 0x25A1, 0x25A3, 0x25A5, 0x25A7, 0x25A9, 0x25AB, 0x25AD, 0x25AF, 0x25B1, 0x25B3, 0x25B5, 0x25B7, 0x25B9, 0x25BB, 0x25BD, 0x25BF, 0x25C1, 0x25C3, 0x25C5, 0x25C7, 0x25C9, 0x25CB, 0x25CD, ( 0x21CF + ( last_character_palette << 10 ) ), 0x201A, 0x2000, 0x2000  ; noqa: E501
-} else {
-    .dw 0x2000, 0x2000, 0x2019, 0x2100, 0x2102, 0x2104, 0x2106, 0x2108, 0x210A, 0x210C, 0x210E, 0x2110, 0x2112, 0x2114, 0x2116, 0x2118, 0x211A, 0x211C, 0x211E, 0x2120, 0x2122, 0x2124, 0x2126, 0x2128, 0x212A, 0x212C, 0x212E, 0x2130, 0x2132, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2101, 0x2103, 0x2105, 0x2107, 0x2109, 0x210B, 0x210D, 0x210F, 0x2111, 0x2113, 0x2115, 0x2117, 0x2119, 0x211B, 0x211D, 0x211F, 0x2121, 0x2123, 0x2125, 0x2127, 0x2129, 0x212B, 0x212D, 0x212F, 0x2131, 0x2133, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2134, 0x2136, 0x2138, 0x213A, 0x213C, 0x213E, 0x2140, 0x2142, 0x2144, 0x2146, 0x2148, 0x214A, 0x214C, 0x214E, 0x2150, 0x2152, 0x2154, 0x2156, 0x2158, 0x215A, 0x215C, 0x215E, 0x2160, 0x2162, 0x2164, 0x2166, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2135, 0x2137, 0x2139, 0x213B, 0x213D, 0x213F, 0x2141, 0x2143, 0x2145, 0x2147, 0x2149, 0x214B, 0x214D, 0x214F, 0x2151, 0x2153, 0x2155, 0x2157, 0x2159, 0x215B, 0x215D, 0x215F, 0x2161, 0x2163, 0x2165, 0x2167, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2168, 0x216A, 0x216C, 0x216E, 0x2170, 0x2172, 0x2174, 0x2176, 0x2178, 0x217A, 0x217C, 0x217E, 0x2180, 0x2182, 0x2184, 0x2186, 0x2188, 0x218A, 0x218C, 0x218E, 0x2190, 0x2192, 0x2194, 0x2196, 0x2198, 0x219A, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x2169, 0x216B, 0x216D, 0x216F, 0x2171, 0x2173, 0x2175, 0x2177, 0x2179, 0x217B, 0x217D, 0x217F, 0x2181, 0x2183, 0x2185, 0x2187, 0x2189, 0x218B, 0x218D, 0x218F, 0x2191, 0x2193, 0x2195, 0x2197, 0x2199, 0x219B, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x219C, 0x219E, 0x21A0, 0x21A2, 0x21A4, 0x21A6, 0x21A8, 0x21AA, 0x21AC, 0x21AE, 0x21B0, 0x21B2, 0x21B4, 0x21B6, 0x21B8, 0x21BA, 0x21BC, 0x21BE, 0x21C0, 0x21C2, 0x21C4, 0x21C6, 0x21C8, 0x21CA, 0x21CC, 0x21CE, 0x201A, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2019, 0x219D, 0x219F, 0x21A1, 0x21A3, 0x21A5, 0x21A7, 0x21A9, 0x21AB, 0x21AD, 0x21AF, 0x21B1, 0x21B3, 0x21B5, 0x21B7, 0x21B9, 0x21BB, 0x21BD, 0x21BF, 0x21C1, 0x21C3, 0x21C5, 0x21C7, 0x21C9, 0x21CB, 0x21CD, 0x21CF, 0x201A, 0x2000, 0x2000  ; noqa: E501
-}
 
-_endwinmap:
-
-_intromap:
-    .dw 0x2000, 0x2000, 0x2000, 0x2100, 0x2102, 0x2104, 0x2106, 0x2108, 0x210A, 0x210C, 0x210E, 0x2110, 0x2112, 0x2114, 0x2116, 0x2118, 0x211A, 0x211C, 0x211E, 0x2120, 0x2122, 0x2124, 0x2126, 0x2128, 0x212A, 0x212C, 0x212E, 0x2130, 0x2132, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x2101, 0x2103, 0x2105, 0x2107, 0x2109, 0x210B, 0x210D, 0x210F, 0x2111, 0x2113, 0x2115, 0x2117, 0x2119, 0x211B, 0x211D, 0x211F, 0x2121, 0x2123, 0x2125, 0x2127, 0x2129, 0x212B, 0x212D, 0x212F, 0x2131, 0x2133, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x2134, 0x2136, 0x2138, 0x213A, 0x213C, 0x213E, 0x2140, 0x2142, 0x2144, 0x2146, 0x2148, 0x214A, 0x214C, 0x214E, 0x2150, 0x2152, 0x2154, 0x2156, 0x2158, 0x215A, 0x215C, 0x215E, 0x2160, 0x2162, 0x2164, 0x2166, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x2135, 0x2137, 0x2139, 0x213B, 0x213D, 0x213F, 0x2141, 0x2143, 0x2145, 0x2147, 0x2149, 0x214B, 0x214D, 0x214F, 0x2151, 0x2153, 0x2155, 0x2157, 0x2159, 0x215B, 0x215D, 0x215F, 0x2161, 0x2163, 0x2165, 0x2167, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x2168, 0x216A, 0x216C, 0x216E, 0x2170, 0x2172, 0x2174, 0x2176, 0x2178, 0x217A, 0x217C, 0x217E, 0x2180, 0x2182, 0x2184, 0x2186, 0x2188, 0x218A, 0x218C, 0x218E, 0x2190, 0x2192, 0x2194, 0x2196, 0x2198, 0x219A, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x2169, 0x216B, 0x216D, 0x216F, 0x2171, 0x2173, 0x2175, 0x2177, 0x2179, 0x217B, 0x217D, 0x217F, 0x2181, 0x2183, 0x2185, 0x2187, 0x2189, 0x218B, 0x218D, 0x218F, 0x2191, 0x2193, 0x2195, 0x2197, 0x2199, 0x219B, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x219C, 0x219E, 0x21A0, 0x21A2, 0x21A4, 0x21A6, 0x21A8, 0x21AA, 0x21AC, 0x21AE, 0x21B0, 0x21B2, 0x21B4, 0x21B6, 0x21B8, 0x21BA, 0x21BC, 0x21BE, 0x21C0, 0x21C2, 0x21C4, 0x21C6, 0x21C8, 0x21CA, 0x21CC, 0x21CE, 0x2000, 0x2000, 0x2000  ; noqa: E501
-    .dw 0x2000, 0x2000, 0x2000, 0x219D, 0x219F, 0x21A1, 0x21A3, 0x21A5, 0x21A7, 0x21A9, 0x21AB, 0x21AD, 0x21AF, 0x21B1, 0x21B3, 0x21B5, 0x21B7, 0x21B9, 0x21BB, 0x21BD, 0x21BF, 0x21C1, 0x21C3, 0x21C5, 0x21C7, 0x21C9, 0x21CB, 0x21CD, 0x21CF, 0x2000, 0x2000, 0x2000  ; noqa: E501
-
-_endintromap:
-
-
-gils_window_size := 10
-
-gils_window_tilemap_1:
-"""Top-edge tilemap row for the gil-count window."""
-    .dw 0, 0
-    .db 0x16, 0x20
-    fill_value(0x2017, gils_window_size - 2)
-    .db 0x18, 0x20
-
-gils_window_tilemap_2:
-"""Filler row for the gil-count window interior."""
-    .dw 0, 0
-
-    .db 0x19, 0x20
-    fill_value(0x20ff, gils_window_size - 2)
-    .db 0x1a, 0x20
-
-gils_window_tilemap_3:
-"""Second filler row for the gil-count window."""
-    .dw 0, 0
-
-    .db 0x19, 0x20
-    fill_value(0x20ff, gils_window_size - 2)
-    .db 0x1a, 0x20
-
-gils_window_tilemap_4:
-"""Gils caption row for the gil-count window."""
-    .dw 0, 0
-
-    .db 0x19, 0x20
-    fill_value(0x20ff, gils_window_size - 2)
-    .db 0x1a, 0x20
-
-    fill_value(0x0000, 0x16)
-
-    .db 0x19, 0x20
-    fill_value(0x20ff, gils_window_size - 2 - 4)
-    .text "G"
-    .db 0x20
-    .text "i"
-    .db 0x20
-    .text "l"
-    .db 0x20
-    .text "s"
-    .db 0x20
-
-    .db 0x1a, 0x20  ; "gil"
-    fill_value(0x0000, 0x16)
-    .db 0x1b, 0x20
-    fill_value(0x201c, gils_window_size - 2)
-    .db 0x1d, 0x20
-
-gils_window_tilemap_4_end:
-"""End-of-data marker for gils_window_tilemap_4."""
+    _endwinmap:
+
+    _intromap:
+        .dw 0x2000, 0x2000, 0x2000, 0x2100, 0x2102, 0x2104, 0x2106, 0x2108, 0x210A, 0x210C, 0x210E, 0x2110, 0x2112, 0x2114, 0x2116, 0x2118, 0x211A, 0x211C, 0x211E, 0x2120, 0x2122, 0x2124, 0x2126, 0x2128, 0x212A, 0x212C, 0x212E, 0x2130, 0x2132, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x2101, 0x2103, 0x2105, 0x2107, 0x2109, 0x210B, 0x210D, 0x210F, 0x2111, 0x2113, 0x2115, 0x2117, 0x2119, 0x211B, 0x211D, 0x211F, 0x2121, 0x2123, 0x2125, 0x2127, 0x2129, 0x212B, 0x212D, 0x212F, 0x2131, 0x2133, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x2134, 0x2136, 0x2138, 0x213A, 0x213C, 0x213E, 0x2140, 0x2142, 0x2144, 0x2146, 0x2148, 0x214A, 0x214C, 0x214E, 0x2150, 0x2152, 0x2154, 0x2156, 0x2158, 0x215A, 0x215C, 0x215E, 0x2160, 0x2162, 0x2164, 0x2166, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x2135, 0x2137, 0x2139, 0x213B, 0x213D, 0x213F, 0x2141, 0x2143, 0x2145, 0x2147, 0x2149, 0x214B, 0x214D, 0x214F, 0x2151, 0x2153, 0x2155, 0x2157, 0x2159, 0x215B, 0x215D, 0x215F, 0x2161, 0x2163, 0x2165, 0x2167, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x2168, 0x216A, 0x216C, 0x216E, 0x2170, 0x2172, 0x2174, 0x2176, 0x2178, 0x217A, 0x217C, 0x217E, 0x2180, 0x2182, 0x2184, 0x2186, 0x2188, 0x218A, 0x218C, 0x218E, 0x2190, 0x2192, 0x2194, 0x2196, 0x2198, 0x219A, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x2169, 0x216B, 0x216D, 0x216F, 0x2171, 0x2173, 0x2175, 0x2177, 0x2179, 0x217B, 0x217D, 0x217F, 0x2181, 0x2183, 0x2185, 0x2187, 0x2189, 0x218B, 0x218D, 0x218F, 0x2191, 0x2193, 0x2195, 0x2197, 0x2199, 0x219B, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x219C, 0x219E, 0x21A0, 0x21A2, 0x21A4, 0x21A6, 0x21A8, 0x21AA, 0x21AC, 0x21AE, 0x21B0, 0x21B2, 0x21B4, 0x21B6, 0x21B8, 0x21BA, 0x21BC, 0x21BE, 0x21C0, 0x21C2, 0x21C4, 0x21C6, 0x21C8, 0x21CA, 0x21CC, 0x21CE, 0x2000, 0x2000, 0x2000  ; noqa: E501
+        .dw 0x2000, 0x2000, 0x2000, 0x219D, 0x219F, 0x21A1, 0x21A3, 0x21A5, 0x21A7, 0x21A9, 0x21AB, 0x21AD, 0x21AF, 0x21B1, 0x21B3, 0x21B5, 0x21B7, 0x21B9, 0x21BB, 0x21BD, 0x21BF, 0x21C1, 0x21C3, 0x21C5, 0x21C7, 0x21C9, 0x21CB, 0x21CD, 0x21CF, 0x2000, 0x2000, 0x2000  ; noqa: E501
+
+    _endintromap:
+
+
+    gils_window_size := 10
+
+    gils_window_tilemap_1:
+    """Top-edge tilemap row for the gil-count window."""
+        .dw 0, 0
+        .db 0x16, 0x20
+        fill_value(0x2017, gils_window_size - 2)
+        .db 0x18, 0x20
+
+    gils_window_tilemap_2:
+    """Filler row for the gil-count window interior."""
+        .dw 0, 0
+
+        .db 0x19, 0x20
+        fill_value(0x20ff, gils_window_size - 2)
+        .db 0x1a, 0x20
+
+    gils_window_tilemap_3:
+    """Second filler row for the gil-count window."""
+        .dw 0, 0
+
+        .db 0x19, 0x20
+        fill_value(0x20ff, gils_window_size - 2)
+        .db 0x1a, 0x20
+
+    gils_window_tilemap_4:
+    """Gils caption row for the gil-count window."""
+        .dw 0, 0
+
+        .db 0x19, 0x20
+        fill_value(0x20ff, gils_window_size - 2)
+        .db 0x1a, 0x20
+
+        fill_value(0x0000, 0x16)
+
+        .db 0x19, 0x20
+        fill_value(0x20ff, gils_window_size - 2 - 4)
+        .text "G"
+        .db 0x20
+        .text "i"
+        .db 0x20
+        .text "l"
+        .db 0x20
+        .text "s"
+        .db 0x20
+
+        .db 0x1a, 0x20  ; "gil"
+        fill_value(0x0000, 0x16)
+        .db 0x1b, 0x20
+        fill_value(0x201c, gils_window_size - 2)
+        .db 0x1d, 0x20
+
+    gils_window_tilemap_4_end:
+    """End-of-data marker for gils_window_tilemap_4."""
+}
diff --git a/src/vwf_state.i b/src/vwf_state.i
new file mode 100644
index 0000000..3786a45
--- /dev/null
+++ b/src/vwf_state.i
@@ -0,0 +1,125 @@
+"""
+Shared VWF engine state addresses + future-config-struct slots.
+
+Both the battle messages renderer (`battle_render` in
+`src/battle/message.s`) and the menu small-VWF renderer
+(`render` in `src/small_vwf/render.s`) build glyph CHR into the same
+SRAM tile buffer at $70:3000. They cohabit by never running
+concurrently: battle owns the buffer during battle scenes, small_vwf
+owns it during field menus.
+
+Define the shared address as a single symbol so future engines
+(field-menu items, treasure-list VWF, drops-list VWF, ...) all
+point at the same place automatically. Bumping the size only
+needs to happen here  ; `_vram_copy.buffer` already sits at $70:5000
+past the buffer so the VRAM-save staging is decoupled from the
+CHR-buffer size.
+
+Forward-compat reservation:
+  VWF_CONFIG_BASE points at a planned per-context state struct
+  (font_ptr, kerning_ptr, tile_id_base, slot_budget, palette,
+  tilemap_base, flags). When the engines start reading from this
+  block, callers will write the struct once before invoking
+  `vwf_render_string` and the engine drops its hardcoded font
+  pointer. The struct lives in unused SRAM space ($70:F800+, same
+  region we already use for the items_menu_vwf debug breadcrumb)
+  so adding it does not perturb battle / menu live state.
+"""
+
+
+; --- VWF CHR buffer (shared across battle + menu renderers) -----------
+
+VWF_CHR_BUFFER := 0x703000
+; Sized to cover tile_ids $00..$1FF (512 * 16 bytes). Field menu
+; rolling buffer has 11 slots * K=10 tiles starting at base $C0,
+; so the high slots end up at tile_id $160+ ; the previous $1000
+; sizing capped at $FF and the high-slot CHR landed in
+; $704000-$704FFF which was outside any blit / DMA reach.
+VWF_CHR_BUFFER_SIZE := 0x2000
+
+; --- Null-terminated text-staging buffer ------------------------------
+; Callers copy the source string (from items_unleashed, monster names,
+; magic list, ...) into this buffer + write $00 terminator, then call
+; `vwf_render_string` with just a pointer. Lets the engine drop the
+; explicit char-count argument that battle / item-description / field
+; helpers each carry today, and lets us swap the source layout
+; (fixed-stride table vs null-terminated table vs RAM-resident string)
+; without touching the renderer. Sized for the longest field-menu
+; item slot in `assets_items_unleashed_dat` + 1 terminator + headroom.
+VWF_TEXT_BUFFER := 0x707000
+VWF_TEXT_BUFFER_SIZE := 0x40
+
+; --- VWF config struct (forward-compat ; not consumed yet) ------------
+; Field offsets so future readers can `lda.l VWF_CONFIG_BASE + .field`
+; without manual byte math. Phase n of the unification will start
+; populating + consuming these.
+VWF_CONFIG_BASE := 0x707080
+
+; --- Engine scratch in SRAM (no DP collisions) ---------------------------
+; Long-addressable scratch for VWF callers that need a counter / pointer
+; without stealing direct-page bytes from the menu loop. The field-items
+; helper uses VWF_SRC_OFFSET as the 16-bit source index into
+; assets_items_unleashed_dat while X holds the destination index in
+; VWF_TEXT_BUFFER (only sta.l abs,x is encoded by a816).
+VWF_SRC_OFFSET := 0x7070C0
+
+; --- Dirty flag for NMI-side CHR flush ----------------------------------
+; Set by `render.display_char` (or `render.render_with_config`) after a
+; blit lands in `VWF_CHR_BUFFER`. The NMI flush hook reads this byte,
+; fires the DMA from `VWF_CHR_BUFFER + $C00` to VRAM $AC00 ($400 bytes)
+; when set, then clears it. Gates the upload exactly the same way the
+; battle inventory's `dma_dirty_slots` gates the per-slot DMA.
+VWF_CHR_DIRTY := 0x7070C2
+
+; --- Secondary descriptor for two-region simultaneous flush -------------
+; Drops + treasure-inventory coexist in the treasure popup and render
+; through the same `items_menu_vwf.draw_field_item_name` JSL hook.
+; Region 1 = $100..$13B (treasure) flushes via the primary descriptor
+; at VWF_CONFIG_BASE ; region 1B = $16E..$1A9 (drops) needs its own
+; flush dest + size so each panel's CHR lands in VRAM without the
+; other's stale buffer bytes leaking through one combined DMA.
+;
+; VWF_CALLER_CTX is a one-byte hint set by drops_rolling around the
+; vanilla JSR chain ; items_menu_vwf reads it to decide whether to
+; write the primary or the secondary descriptor + dirty flag.
+;   0 = primary  (treasure / field-items / default)
+;   1 = drops    (secondary)
+VWF_CALLER_CTX := 0x7070C3
+VWF_CHR_DIRTY_B := 0x7070C4
+VWF_CHR_VRAM_WORD_B := 0x7070C5
+VWF_CHR_BYTE_COUNT_B := 0x7070C7
+VWF_CHR_SRC_OFFSET_B := 0x7070C9
+
+.struct VwfConfig {
+    word tile_id_base
+    byte slot_budget
+    word tilemap_base
+    byte palette_byte
+    byte flags
+    word chr_vram_word
+    word chr_byte_count
+}
+
+; Engine-shared CHR-flush source offset. Every VWF caller writes
+; glyph CHR at `VWF_CHR_BUFFER + tile_id_base * 16` ; both battle
+; inventory and field menu items happen to anchor at tile_id base
+; $C0, so the flush always starts $C00 bytes into the buffer.
+; (If a future client uses a different base, this constant moves to
+; the engine and we still avoid forking the DMA source per caller.)
+VWF_CHR_FLUSH_OFFSET := 0x100 * 0x10
+
+; --- VRAM save / restore range ---
+; The dialog VWF saves the VRAM CHR window before dialog opens and
+; restores it on close. With field-menu items + item-description
+; each carving out their own slice of CHR ($5000..$5FFF), the save
+; range must cover that whole 4KB block so menu close puts the
+; menu CHR back the way the field renderer expects it.
+;
+;   Word $2800 .. $37FF  =  byte $5000 .. $6FFF  (8KB)
+;
+; Generous on the upper end so a future region 3 / region 4 (status,
+; equipment, ...) lands inside the save window without bumping the
+; SRAM scratch.
+VRAM_SAVE_BASE_WORD := 0x2800
+VRAM_SAVE_BYTE_COUNT := 0x2000
+VRAM_SAVE_SRAM_BASE := 0x705000
diff --git a/tests/_profile/bench_nmi.py b/tests/_profile/bench_nmi.py
index 45a0edb..38fbc6e 100644
--- a/tests/_profile/bench_nmi.py
+++ b/tests/_profile/bench_nmi.py
@@ -19,7 +19,7 @@
     0x028302: "SetFlyingHDMA (added NMI hook)",
 }
 
-e = load_emu_from_kss(kss_path("ff4-battle-ext.kss"))
+e = load_emu_from_kss(kss_path("ff4-battle.kss"))
 e.run_frames(60)
 
 e.profile_start(lo=0x000000, hi=0xFFFFFF)
diff --git a/tests/_profile/dump_inv_sram.py b/tests/_profile/dump_inv_sram.py
new file mode 100644
index 0000000..b7383cc
--- /dev/null
+++ b/tests/_profile/dump_inv_sram.py
@@ -0,0 +1,54 @@
+"""Dump WRAM tile buffer + inventory slot tilemap-buffer to text for
+visual inspection. Loads battle-ext, opens inventory, snapshots."""
+from __future__ import annotations
+import sys
+from pathlib import Path
+
+sys.path.insert(0, "tests")
+from kintsuki import Button
+from _ff4kintsuki import kss_path, load_emu_from_kss, tap
+
+OUT = Path(__file__).parent / "dumps" / "inv_sram.hex"
+OUT.parent.mkdir(exist_ok=True)
+
+
+def dump_region(emu, label: str, start: int, length: int) -> str:
+    rows = []
+    rows.append(f"=== {label}  ${start:06x}..${start+length-1:06x} ({length}B) ===")
+    for off in range(0, length, 16):
+        addr = start + off
+        bytes_ = " ".join(f"{emu.read(addr + i):02x}" for i in range(16))
+        ascii_ = "".join(
+            chr(b) if 32 <= b < 127 else "."
+            for b in (emu.read(addr + i) for i in range(16))
+        )
+        rows.append(f"{addr:06x}: {bytes_}  {ascii_}")
+    return "\n".join(rows)
+
+
+def main() -> None:
+    emu = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=600)
+    tap(emu, Button.DOWN, gap=20)
+    tap(emu, Button.DOWN, gap=20)
+    tap(emu, Button.A, gap=60)
+
+    out = []
+    # Inventory rolling-buffer tilemap entries (6 slots * 60 bytes)
+    out.append(dump_region(emu, "rolling slots ($7E97A6)", 0x7E97A6, 6 * 60))
+    # VWF tile CHR staging
+    out.append(dump_region(emu, "vwf buffer head (msg/mon/names)", 0x703000, 0x600))
+    out.append(dump_region(emu, "vwf buffer mid (cmd)", 0x703900, 0x300))
+    out.append(dump_region(emu, "vwf buffer tail (inventory)", 0x703C00, 0x300))
+    out.append(dump_region(emu, "vwf gate state", 0x703F00, 0x10))
+    # Allocator state
+    out.append(dump_region(emu, "allocator", 0x702F00, 0x10))
+    # Format buffer used by inventory render
+    out.append(dump_region(emu, "inv_format_buffer ($7E9E66)", 0x7E9E66, 0x40))
+    # Mode flag DP $37
+    out.append(dump_region(emu, "mode flag $37", 0x7E0030, 0x10))
+    OUT.write_text("\n\n".join(out))
+    print(f"wrote {OUT} ({OUT.stat().st_size} bytes)")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/_profile/dump_inv_tilemap.py b/tests/_profile/dump_inv_tilemap.py
new file mode 100644
index 0000000..52305d6
--- /dev/null
+++ b/tests/_profile/dump_inv_tilemap.py
@@ -0,0 +1,122 @@
+"""Snapshot the battle inventory rolling-buffer tilemap + allocator
+state across consecutive frames. Diffs each frame's snapshot against
+the prior frame so we can pin down whether the tile_id refs in the
+tilemap drift (allocator races), whether the WRAM CHR buffer drifts
+(buffer reuse race), or both stay stable while VRAM stays correct.
+
+Usage:
+    python3 tests/_profile/dump_inv_tilemap.py
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+sys.path.insert(0, "tests")
+from kintsuki import Button
+from _ff4kintsuki import kss_path, load_emu_from_kss, tap
+
+REPO = Path(__file__).resolve().parents[2]
+SYM_PATH = REPO / "build" / "ff4.sym"
+
+# Inventory rolling-buffer tilemap entries live in WRAM around the
+# vanilla DrawInventoryItemText slot. The rolling layout stores 6
+# slots * 60 bytes starting at $97A6.
+SLOT_BASE = 0x7E97A6
+SLOT_STRIDE = 60
+SLOT_COUNT = 6
+
+# Inventory CHR region in the shared VWF buffer (16 bytes per tile,
+# tile_id 0xC0..0xEF -> buffer + $C00..$EF0).
+BUFFER_CHR_BASE = 0x703C00
+BUFFER_CHR_LEN = 0x300  # 48 tiles * 16
+
+# Allocator state (cross-bank).
+ALLOCATED_TILE_ID = 0x702F00
+
+# Gate state (recently moved past inventory tiles).
+PENDING_TRANSFER_MASK = 0x703F00
+REGION_DIRTY_BITS = 0x703F01
+RENDER_SKIPPED = 0x703F02
+TILEMAP_PENDING_MASK = 0x703F03
+
+# Rolling-buffer state.
+ROLLING_TOP_ROW = 0x7EEF84
+ROLLING_BUFFER_POS = 0x7EEF80
+ROLLING_SLOT_INDEX = 0x7EEF82
+
+
+def snapshot(emu) -> dict:
+    snap = {
+        "tile_id": emu.read(ALLOCATED_TILE_ID),
+        "pending": emu.read(PENDING_TRANSFER_MASK),
+        "region_dirty": emu.read(REGION_DIRTY_BITS),
+        "render_skipped": emu.read(RENDER_SKIPPED),
+        "tilemap_pending": emu.read(TILEMAP_PENDING_MASK),
+        "rolling_top": emu.read(ROLLING_TOP_ROW),
+        "rolling_pos": emu.read(ROLLING_BUFFER_POS),
+        "rolling_slot": emu.read(ROLLING_SLOT_INDEX),
+        "slots": [
+            bytes(emu.read(SLOT_BASE + i * SLOT_STRIDE + j) for j in range(SLOT_STRIDE))
+            for i in range(SLOT_COUNT)
+        ],
+        "buffer_chr": bytes(emu.read(BUFFER_CHR_BASE + i) for i in range(BUFFER_CHR_LEN)),
+    }
+    return snap
+
+
+def diff(prev: dict, cur: dict, frame: int) -> None:
+    scalars = [
+        "tile_id", "pending", "region_dirty", "render_skipped",
+        "tilemap_pending", "rolling_top", "rolling_pos", "rolling_slot",
+    ]
+    changes = []
+    for k in scalars:
+        if prev[k] != cur[k]:
+            changes.append(f"{k}: {prev[k]:02X} -> {cur[k]:02X}")
+    for i in range(SLOT_COUNT):
+        if prev["slots"][i] != cur["slots"][i]:
+            differing = sum(1 for a, b in zip(prev["slots"][i], cur["slots"][i]) if a != b)
+            changes.append(f"slot{i}: {differing} bytes differ")
+    if prev["buffer_chr"] != cur["buffer_chr"]:
+        differing = sum(1 for a, b in zip(prev["buffer_chr"], cur["buffer_chr"]) if a != b)
+        changes.append(f"buffer_chr: {differing} / {BUFFER_CHR_LEN} bytes differ")
+    print(f"[frame {frame}] " + (", ".join(changes) if changes else "no change"))
+
+
+def main() -> None:
+    emu = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=600)
+    # Settle until Cecil's ATB fills + cmd menu opens, then drive the
+    # input sequence the user mapped out: DOWN, DOWN, A -> item menu.
+    tap(emu, Button.DOWN, gap=20)
+    tap(emu, Button.DOWN, gap=20)
+    tap(emu, Button.A, gap=60)
+    print(f"Initial state after settle + nav:")
+    s0 = snapshot(emu)
+    print(f"  tile_id={s0['tile_id']:02X} pending={s0['pending']:02X} "
+          f"region_dirty={s0['region_dirty']:02X} rolling_slot={s0['rolling_slot']:02X}")
+    for i, slot in enumerate(s0["slots"]):
+        first16 = slot[:16].hex(" ")
+        print(f"  slot{i}[:16]: {first16}")
+    print()
+
+    prev = s0
+    for f in range(1, 11):
+        emu.run_frames(1)
+        cur = snapshot(emu)
+        diff(prev, cur, f)
+        prev = cur
+
+    # Capture the visible framebuffer so we can eyeball the "scrolling"
+    # symptom alongside the byte deltas.
+    from kintsuki.visual import golden
+    out = REPO / "tests" / "goldens" / "battle_init" / "inv_open.png"
+    out.parent.mkdir(parents=True, exist_ok=True)
+    if out.exists():
+        out.unlink()
+    golden(emu, out, threshold=0.1, max_diff_pixels=0)
+    print(f"\nframebuffer saved to {out}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tests/_profile/dumps/inv_sram.hex b/tests/_profile/dumps/inv_sram.hex
new file mode 100644
index 0000000..fb280c0
--- /dev/null
+++ b/tests/_profile/dumps/inv_sram.hex
@@ -0,0 +1,237 @@
+=== rolling slots ($7E97A6)  $7e97a6..$7e990d (360B) ===
+7e97a6: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7e97b6: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7e97c6: c0 01 c1 01 c2 01 c3 01 c4 01 c5 01 ff 00 ff 00  ................
+7e97d6: ff 00 ff 00 ff 00 c8 00 80 00 85 00 ff 00 ff 00  ................
+7e97e6: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7e97f6: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ca 01 cb 01  ................
+7e9806: cc 01 cd 01 ce 01 cf 01 ff 00 ff 00 ff 00 ff 00  ................
+7e9816: ff 00 c8 00 80 00 86 00 ff 04 ff 04 ff 04 ff 04  ................
+7e9826: ff 04 ff 04 ff 04 ff 04 ff 00 ff 00 ff 00 ff 00  ................
+7e9836: ff 04 ff 04 ff 04 2a 04 d4 05 d5 05 d6 05 d7 05  ......*.........
+7e9846: d8 05 d9 05 da 05 ff 00 ff 00 ff 00 ff 00 c8 04  ................
+7e9856: ff 04 81 04 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7e9866: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7e9876: ff 00 ff 00 de 01 df 01 e0 01 e1 01 e2 01 e3 01  ................
+7e9886: ff 00 ff 00 ff 00 ff 00 ff 00 e3 01 e4 01 ff 00  ................
+7e9896: ff 04 ff 04 ff 04 ff 04 ff 04 ff 04 ff 04 ff 04  ................
+7e98a6: ff 00 ff 00 ff 00 ff 00 ff 04 ff 04 ff 04 2a 04  ..............*.
+7e98b6: e8 05 e9 05 ea 05 eb 05 ec 05 ed 05 ee 05 ff 00  ................
+7e98c6: ff 00 ff 00 ff 00 c8 04 ff 04 81 04 ff 00 ff 00  ................
+7e98d6: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7e98e6: ff 00 ff 00 ff 00 ff 00 ff 00 2a 00 f2 01 f3 01  ..........*.....
+7e98f6: f4 01 f5 01 f6 01 f7 01 f8 01 ff 00 ff 00 ff 00  ................
+7e9906: ff 00 c8 00 ff 00 81 00 00 00 00 00 00 00 00 00  ................
+
+=== vwf buffer head (msg/mon/names)  $703000..$7035ff (1536B) ===
+703000: ff 00 ff 62 ff 92 ff 97 ff f2 ff 92 ff 92 ff 91  ...b............
+703010: ff 00 ff 20 ff 20 ff 76 ff 21 ff 27 ff 29 ff 17  ... . .v.!.'.)..
+703020: ff 00 ff 00 ff 00 ff 32 ff 4a ff 4a ff 3a ff 09  .......2.J.J.:..
+703030: ff 00 ff 00 ff 00 ff 4c ff 52 ff 5e ff 50 ff 8c  .......L.R.^.P..
+703040: ff 00 ff 00 ff 00 ff 0e ff 09 ff 09 ff 0e ff 08  ................
+703050: ff 01 ff 02 ff 00 ff 53 ff 64 ff 47 ff 44 ff 43  .......S.d.G.D.C
+703060: ff 00 ff 00 ff 00 ff 24 ff a5 ff a9 ff 31 ff 20  .......$.....1. 
+703070: ff 00 ff 00 ff 00 ff ce ff 29 ff e9 ff 09 ff c9  .........)......
+703080: ff 00 ff 48 ff 40 ff ea ff 4a ff 4a ff 4b ff 2a  ...H.@...J.J.K.*
+703090: ff 00 ff 00 ff 00 ff 4c ff 52 ff 9e ff 10 ff 0c  .......L.R......
+7030a0: ff 00 ff 08 ff 08 ff 08 ff 08 ff 08 ff 00 ff 08  ................
+7030b0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7030c0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7030d0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7030e0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7030f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703100: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703110: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703120: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703130: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703140: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703150: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703160: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703170: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703180: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703190: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7031a0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7031b0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7031c0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7031d0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7031e0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7031f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703200: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703210: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703220: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703230: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703240: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703250: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703260: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703270: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703280: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703290: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7032a0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7032b0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7032c0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7032d0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7032e0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7032f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703300: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703310: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703320: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703330: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703340: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703350: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703360: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703370: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703380: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703390: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7033a0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7033b0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7033c0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7033d0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7033e0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7033f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703400: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703410: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703420: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703430: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703440: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703450: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703460: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703470: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703480: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703490: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7034a0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7034b0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7034c0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7034d0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7034e0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7034f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703500: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703510: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703520: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703530: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703540: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703550: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703560: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703570: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703580: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703590: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7035a0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7035b0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7035c0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7035d0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7035e0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7035f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+
+=== vwf buffer mid (cmd)  $703900..$703bff (768B) ===
+703900: ff 00 ff 62 ff 92 ff 97 ff f2 ff 92 ff 92 ff 91  ...b............
+703910: ff 00 ff 20 ff 20 ff 76 ff 21 ff 27 ff 29 ff 17  ... . .v.!.'.)..
+703920: ff 00 ff 00 ff 00 ff 32 ff 4a ff 4a ff 3a ff 09  .......2.J.J.:..
+703930: ff 00 ff 00 ff 00 ff 4c ff 52 ff 5e ff 50 ff 8c  .......L.R.^.P..
+703940: ff 00 ff 00 ff 00 ff a0 ff c0 ff 80 ff 80 ff 80  ................
+703950: ff 00 ff 64 ff 94 ff 97 ff 94 ff 94 ff 94 ff 67  ...d...........g
+703960: ff 00 ff 00 ff 00 ff 18 ff a1 ff 99 ff 85 ff 38  ...............8
+703970: ff 00 ff 00 ff 00 ff c9 ff 29 ff 09 ff 29 ff c6  .........)...)..
+703980: ff 00 ff 04 ff 00 ff 55 ff 64 ff 44 ff 44 ff 44  .......U.d.D.D.D
+703990: ff 04 ff 88 ff 80 ff cc ff 92 ff 9e ff 90 ff 4c  ...............L
+7039a0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+7039b0: ff 00 ff 64 ff 94 ff 97 ff 94 ff 94 ff 94 ff 67  ...d...........g
+7039c0: ff 00 ff 10 ff 00 ff 13 ff 94 ff 97 ff 94 ff 23  ...............#
+7039d0: ff 00 ff 10 ff 10 ff 39 ff 92 ff 91 ff 10 ff 0b  .......9........
+7039e0: ff 00 ff 00 ff 00 ff 80 ff 00 ff 80 ff 40 ff 80  .............@..
+7039f0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a00: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a10: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a20: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a30: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a40: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a50: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a60: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a70: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a80: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703a90: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703aa0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ab0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ac0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ad0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ae0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703af0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b00: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b10: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b20: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b30: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b40: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b50: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b60: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b70: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b80: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703b90: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ba0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703bb0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703bc0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703bd0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703be0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703bf0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+
+=== vwf buffer tail (inventory)  $703c00..$703eff (768B) ===
+703c00: ff 00 ff 64 ff 90 ff 94 ff f5 ff 94 ff 94 ff 95  ...d............
+703c10: ff 00 ff 00 ff 00 ff e9 ff 29 ff e9 ff 29 ff c6  .........)...)..
+703c20: ff 00 ff 54 ff 14 ff 54 ff 55 ff 55 ff 55 ff 54  ...T...T.U.U.U.T
+703c30: ff 00 ff 00 ff 00 ff c0 ff 20 ff e0 ff 00 ff c0  ......... ......
+703c40: ff 00 ff 60 ff 90 ff 95 ff 96 ff 94 ff 94 ff 64  ...`...........d
+703c50: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703c60: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703c70: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703c80: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703c90: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+703ca0: ff 00 ff f2 ff 82 ff 87 ff e2 ff 82 ff 82 ff f1  ................
+703cb0: ff 00 ff 40 ff 40 ff 71 ff 4a ff 4b ff 4a ff 49  ...@.@.q.J.K.J.I
+703cc0: ff 00 ff 00 ff 00 ff 94 ff 58 ff d0 ff 10 ff 90  .........X......
+703cd0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ce0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703cf0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703d00: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703d10: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703d20: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703d30: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+703d40: ff 00 ff e0 ff 90 ff 93 ff e0 ff 93 ff 94 ff e3  ................
+703d50: ff 00 ff 00 ff 00 ff 1d ff a5 ff 9d ff 85 ff b8  ................
+703d60: ff 00 ff 00 ff 00 ff 26 ff 29 ff 2f ff 28 ff c6  .......&.)./.(..
+703d70: ff 00 ff 22 ff 22 ff 77 ff 22 ff 22 ff 22 ff 11  ...".".w."."."..
+703d80: ff 00 ff 00 ff 00 ff 30 ff 48 ff 78 ff 40 ff 30  .......0.H.x.@.0
+703d90: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703da0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703db0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703dc0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703dd0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+703de0: ff 00 ff 60 ff 90 ff 83 ff 84 ff 84 ff 94 ff 63  ...`...........c
+703df0: ff 00 ff 28 ff 28 ff 2a ff aa ff a9 ff a8 ff 2b  ...(.(.*.......+
+703e00: ff 00 ff 00 ff 00 ff 55 ff 5a ff d3 ff 52 ff 91  .......U.Z...R..
+703e10: ff 00 ff 00 ff 00 ff 80 ff 40 ff c0 ff 00 ff 80  .........@......
+703e20: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703e30: ff 00 ff 18 ff 21 ff a0 ff 38 ff 24 ff a4 ff 19  .....!...8.$....
+703e40: ff 00 ff 80 ff 80 ff 80 ff 80 ff 80 ff 80 ff c0  ................
+703e50: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703e60: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703e70: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+703e80: ff 00 ff e0 ff 90 ff 93 ff e0 ff 93 ff 94 ff e3  ................
+703e90: ff 00 ff 00 ff 00 ff 1d ff a5 ff 9d ff 85 ff b8  ................
+703ea0: ff 00 ff 00 ff 00 ff 26 ff 29 ff 2f ff 28 ff c6  .......&.)./.(..
+703eb0: ff 00 ff 22 ff 22 ff 77 ff 22 ff 22 ff 22 ff 11  ...".".w."."."..
+703ec0: ff 00 ff 00 ff 00 ff 30 ff 48 ff 78 ff 40 ff 30  .......0.H.x.@.0
+703ed0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ee0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+703ef0: ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+
+=== vwf gate state  $703f00..$703f0f (16B) ===
+703f00: e8 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00 ff 00  ................
+
+=== allocator  $702f00..$702f0f (16B) ===
+702f00: ee 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+
+=== inv_format_buffer ($7E9E66)  $7e9e66..$7e9ea5 (64B) ===
+7e9e66: 0e 04 2a fe 0e 04 43 5c 62 70 60 6f 6f 60 ff ff  ..*...C\bp`oo`..
+7e9e76: ff fe fc 0c c8 ff 81 00 00 00 00 00 00 00 00 00  ................
+7e9e86: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+7e9e96: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
+
+=== mode flag $37  $7e0030..$7e003f (16B) ===
+7e0030: 05 75 10 d8 50 d8 20 00 00 00 00 00 00 00 00 c2  .u..P. .........
\ No newline at end of file
diff --git a/tests/_profile/profile_battle_vwf.py b/tests/_profile/profile_battle_vwf.py
index b403fc4..530189e 100644
--- a/tests/_profile/profile_battle_vwf.py
+++ b/tests/_profile/profile_battle_vwf.py
@@ -1,6 +1,6 @@
 """Profile the battle VWF blit path with all redraw gates forced dirty.
 
-Loads ff4-battle-ext.kss at the Battle_ext seed point, pokes every
+Loads ff4-battle.kss at the Battle_ext seed point, pokes every
 redraw-gate dirty bit so DrawCharNames + DrawMonsterNames + DrawCmd
 fire on every profiled frame, then runs `profile_start` scoped to the
 display_char neighbourhood in bank $20 and reports per-function call
@@ -54,11 +54,11 @@ def label(pc: int) -> str:
 def force_redraw(emu):
     emu.write(0x7EEF9A, 0xFF)  # battle_menu_dirty: chars + cmd + status
     emu.write(0x7EEF9B, 0xFF)  # battle_monster_dirty: all monster slots
-    emu.write(0x703C01, 0xFF)  # region_dirty_bits: slice-2 queue
+    emu.write(0x703F01, 0xFF)  # region_dirty_bits: slice-2 queue (moved past inventory tiles)
 
 
 def run():
-    emu = load_emu_from_kss(kss_path("ff4-battle-ext.kss"), settle_frames=0)
+    emu = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=0)
     force_redraw(emu)
     emu.run_frames(WARMUP_FRAMES)
     force_redraw(emu)
diff --git a/tests/_profile/trace_cmd_dma.py b/tests/_profile/trace_cmd_dma.py
index 8ee789e..e26eaa4 100644
--- a/tests/_profile/trace_cmd_dma.py
+++ b/tests/_profile/trace_cmd_dma.py
@@ -3,7 +3,7 @@
 sys.path.insert(0, "tests")
 from _ff4kintsuki import kss_path, load_emu_from_kss
 
-e = load_emu_from_kss(kss_path("ff4-battle-ext.kss"), settle_frames=0)
+e = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=0)
 
 def snap(label):
     cmd_dirty = e.read(0x7EEF9A)
diff --git a/tests/_profile/trace_drift.py b/tests/_profile/trace_drift.py
index 07a030b..baa82e8 100644
--- a/tests/_profile/trace_drift.py
+++ b/tests/_profile/trace_drift.py
@@ -2,11 +2,10 @@
 import sys
 sys.path.insert(0, "tests")
 import ctypes
-from pathlib import Path
 from kintsuki import Emu
-from _ff4kintsuki import ROM
+from _ff4kintsuki import ROM, kss_path
 
-KSS = Path("/Users/manz/PyCharmProjects/ff4-modules/.claude/worktrees/equip-window-relayout/tests/savestates/ff4-battle-ext.kss")
+KSS = kss_path("ff4-battle.kss")
 e = Emu(load_srm_sidecar=False)
 e.load_rom(str(ROM))
 blob = KSS.read_bytes()
diff --git a/tests/_profile/trace_flying.py b/tests/_profile/trace_flying.py
index 3246453..57a985e 100644
--- a/tests/_profile/trace_flying.py
+++ b/tests/_profile/trace_flying.py
@@ -3,7 +3,7 @@
 sys.path.insert(0, "tests")
 from _ff4kintsuki import kss_path, load_emu_from_kss
 
-e = load_emu_from_kss(kss_path("ff4-battle-ext.kss"), settle_frames=0)
+e = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=0)
 
 def snap(label):
     ed4e = e.read(0x7EED4E)
diff --git a/tests/_profile/trace_init_highlight.py b/tests/_profile/trace_init_highlight.py
index 03a4a71..11e4f1c 100644
--- a/tests/_profile/trace_init_highlight.py
+++ b/tests/_profile/trace_init_highlight.py
@@ -23,7 +23,7 @@ def slot_pal_bytes(slot):
 def snap(label):
     s1822 = e.read(0x7E1822)
     dirty = e.read(0x7EEF9A)
-    region = e.read(0x703C01)
+    region = e.read(0x703F01)
     pals = [slot_pal_bytes(s) for s in range(5)]
     print(f"[{label:>10}] $1822={s1822:02X} EF9A={dirty:02X} reg_dirty={region:02X}")
     for s, pal in enumerate(pals):
diff --git a/tests/_profile/trace_names_layout.py b/tests/_profile/trace_names_layout.py
index 26c1869..5346db6 100644
--- a/tests/_profile/trace_names_layout.py
+++ b/tests/_profile/trace_names_layout.py
@@ -3,7 +3,7 @@
 sys.path.insert(0, "tests")
 from _ff4kintsuki import kss_path, load_emu_from_kss
 
-e = load_emu_from_kss(kss_path("ff4-battle-ext.kss"), settle_frames=0)
+e = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=0)
 e.run_frames(120)
 
 # $74FD = text buffer (input to VWF), built by UpdateCharNames
diff --git a/tests/_profile/trace_palette.py b/tests/_profile/trace_palette.py
index 3b3269a..a98afa8 100644
--- a/tests/_profile/trace_palette.py
+++ b/tests/_profile/trace_palette.py
@@ -3,7 +3,7 @@
 sys.path.insert(0, "tests")
 from _ff4kintsuki import kss_path, load_emu_from_kss
 
-e = load_emu_from_kss(kss_path("ff4-battle-ext.kss"), settle_frames=0)
+e = load_emu_from_kss(kss_path("ff4-battle.kss"), settle_frames=0)
 e.run_frames(120)
 
 def dump_b966():
diff --git a/tests/goldens/battle_init/battle_init.png b/tests/goldens/battle_init/battle_init.png
index c8f0955..b1b38d0 100644
Binary files a/tests/goldens/battle_init/battle_init.png and b/tests/goldens/battle_init/battle_init.png differ
diff --git a/tests/goldens/battle_init/inv_open.png b/tests/goldens/battle_init/inv_open.png
new file mode 100644
index 0000000..0b2d8aa
Binary files /dev/null and b/tests/goldens/battle_init/inv_open.png differ
diff --git a/tests/goldens/field_inventory/scroll_0.bin b/tests/goldens/field_inventory/scroll_0.bin
index 858e02a..de37067 100644
Binary files a/tests/goldens/field_inventory/scroll_0.bin and b/tests/goldens/field_inventory/scroll_0.bin differ
diff --git a/tests/goldens/field_inventory/scroll_0.png b/tests/goldens/field_inventory/scroll_0.png
index 15d9d56..9b7459d 100644
Binary files a/tests/goldens/field_inventory/scroll_0.png and b/tests/goldens/field_inventory/scroll_0.png differ
diff --git a/tests/goldens/field_inventory/scroll_1.bin b/tests/goldens/field_inventory/scroll_1.bin
index 858e02a..de37067 100644
Binary files a/tests/goldens/field_inventory/scroll_1.bin and b/tests/goldens/field_inventory/scroll_1.bin differ
diff --git a/tests/goldens/field_inventory/scroll_10.bin b/tests/goldens/field_inventory/scroll_10.bin
index f33dd9e..6cff6a8 100644
Binary files a/tests/goldens/field_inventory/scroll_10.bin and b/tests/goldens/field_inventory/scroll_10.bin differ
diff --git a/tests/goldens/field_inventory/scroll_10.png b/tests/goldens/field_inventory/scroll_10.png
index dcb058e..7cd27b3 100644
Binary files a/tests/goldens/field_inventory/scroll_10.png and b/tests/goldens/field_inventory/scroll_10.png differ
diff --git a/tests/goldens/field_inventory/scroll_15.bin b/tests/goldens/field_inventory/scroll_15.bin
index 6ed5e8e..f61c3ee 100644
Binary files a/tests/goldens/field_inventory/scroll_15.bin and b/tests/goldens/field_inventory/scroll_15.bin differ
diff --git a/tests/goldens/field_inventory/scroll_5.bin b/tests/goldens/field_inventory/scroll_5.bin
index 858e02a..de37067 100644
Binary files a/tests/goldens/field_inventory/scroll_5.bin and b/tests/goldens/field_inventory/scroll_5.bin differ
diff --git a/tests/goldens/field_inventory/scroll_5.png b/tests/goldens/field_inventory/scroll_5.png
index 6f9521f..180d38d 100644
Binary files a/tests/goldens/field_inventory/scroll_5.png and b/tests/goldens/field_inventory/scroll_5.png differ
diff --git a/tests/goldens/field_inventory/swap_0_to_10.bg1.bin b/tests/goldens/field_inventory/swap_0_to_10.bg1.bin
index 344de30..3cba4e3 100644
Binary files a/tests/goldens/field_inventory/swap_0_to_10.bg1.bin and b/tests/goldens/field_inventory/swap_0_to_10.bg1.bin differ
diff --git a/tests/goldens/field_inventory/swap_0_to_3.bg1.bin b/tests/goldens/field_inventory/swap_0_to_3.bg1.bin
index 70989a3..1057f3a 100644
Binary files a/tests/goldens/field_inventory/swap_0_to_3.bg1.bin and b/tests/goldens/field_inventory/swap_0_to_3.bg1.bin differ
diff --git a/tests/goldens/field_inventory/trash_at_row3.bin b/tests/goldens/field_inventory/trash_at_row3.bin
index 858e02a..de37067 100644
Binary files a/tests/goldens/field_inventory/trash_at_row3.bin and b/tests/goldens/field_inventory/trash_at_row3.bin differ
diff --git a/tests/savestates/ff4-before-field-inventory.kss b/tests/savestates/ff4-before-field-inventory.kss
new file mode 100644
index 0000000..12b863b
Binary files /dev/null and b/tests/savestates/ff4-before-field-inventory.kss differ
diff --git a/tests/savestates/ff4-battle-ext.kss b/tests/savestates/ff4-before-inventory-opens.kss
similarity index 65%
rename from tests/savestates/ff4-battle-ext.kss
rename to tests/savestates/ff4-before-inventory-opens.kss
index 28f8fae..40ae7e6 100644
Binary files a/tests/savestates/ff4-battle-ext.kss and b/tests/savestates/ff4-before-inventory-opens.kss differ
diff --git a/tests/savestates/ff4-before-mayhem2.kss b/tests/savestates/ff4-before-mayhem2.kss
new file mode 100644
index 0000000..7c95131
Binary files /dev/null and b/tests/savestates/ff4-before-mayhem2.kss differ
diff --git a/tests/savestates/ff4-before-treasure-mayhem.kss b/tests/savestates/ff4-before-treasure-mayhem.kss
new file mode 100644
index 0000000..16293ee
Binary files /dev/null and b/tests/savestates/ff4-before-treasure-mayhem.kss differ
diff --git a/tests/test_battle_init.py b/tests/test_battle_init.py
index a8765e4..1be873a 100644
--- a/tests/test_battle_init.py
+++ b/tests/test_battle_init.py
@@ -24,7 +24,7 @@
 )
 
 GOLDENS = Path(__file__).parent / "goldens" / "battle_init"
-KSS = kss_path("ff4-battle-ext.kss")
+KSS = kss_path("ff4-battle.kss")
 
 
 @pytest.fixture
@@ -37,7 +37,7 @@ def battle_emu():
     # or HP delta) and the golden captures an empty header strip.
     emu.write(0x7EEF9A, 0xFF)  # battle_menu_dirty: all chars + cmd + status
     emu.write(0x7EEF9B, 0xFF)  # battle_monster_dirty: all monster slots
-    emu.write(0x703C01, 0xFF)  # region_dirty_bits: slice-2 queue
+    emu.write(0x703F01, 0xFF)  # region_dirty_bits: slice-2 queue (moved past inventory tiles)
     emu.run_frames(600)
     yield emu
     emu.close()
diff --git a/tests/test_field_inventory_rolling.py b/tests/test_field_inventory_rolling.py
index 7464005..9d71121 100644
--- a/tests/test_field_inventory_rolling.py
+++ b/tests/test_field_inventory_rolling.py
@@ -40,14 +40,28 @@
 )
 
 GOLDENS = Path(__file__).parent / "goldens" / "field_inventory"
-KSS = kss_path("ff4-field-inventory-open.kss")
+# Pre-open kss : the legacy `ff4-field-inventory-open.kss` froze
+# VRAM from before the items_menu_vwf flush split + bottom-row blank
+# pre-fill landed, so loading it short-circuits the new render path
+# and the post-fix CHR never reaches VRAM. Use the before-open state
+# and drive a fresh open through the current ROM so every render hook
+# lands the way it does in-game.
+KSS = kss_path("ff4-before-inventory-opens.kss")
 
 
 @pytest.fixture
 def field_emu():
-    e = load_emu_from_kss(KSS)
-    # Settle so the menu is fully open and HDMA shadows are populated.
+    e = load_emu_from_kss(KSS, settle_frames=60)
+    # Open the items submenu.
+    tap(e, Button.A)
     e.run_frames(60)
+    # Cursor wiggle nudges check_if_description_was_rendered off its
+    # initial "last_drawn == 0 / autocounter pending" state so the
+    # description redraw actually fires for the highlighted slot.
+    tap(e, Button.DOWN)
+    e.run_frames(15)
+    tap(e, Button.UP)
+    e.run_frames(120)
     yield e
     e.close()
 
diff --git a/tests/test_menu_exit_cleanup.py b/tests/test_menu_exit_cleanup.py
index b14af0c..553d1d8 100644
--- a/tests/test_menu_exit_cleanup.py
+++ b/tests/test_menu_exit_cleanup.py
@@ -14,7 +14,7 @@
 )
 
 FIELD_STATE_BASE = 0x7E1BA8
-TREASURE_STATE_BASE = 0x7E1BD0
+TREASURE_STATE_BASE = 0x7E9C00
 HDMA_SHADOW = 0x7E1BAE
 TREASURE_IN_MENU_FLAG = 0x7E1BC6
 HDMA5_BASE = 0x004350
diff --git a/tests/test_rolling_engine_integration.py b/tests/test_rolling_engine_integration.py
new file mode 100644
index 0000000..c73fbf2
--- /dev/null
+++ b/tests/test_rolling_engine_integration.py
@@ -0,0 +1,115 @@
+"""Integration test : drive the bank-20 rolling-inventory engine against
+real in-game state, not synthetic SCRATCH_STATE.
+
+Opens the field-menu items submenu (state base = $7E:1BA8) and the
+treasure popup state (base = $7E:1BD0), then JSLs through
+`rolling_engine_invalidate_all` + `rolling_engine_vblank_flush` to
+verify the dirty_mask gate works on live WRAM.
+
+Phase 1 entries are profile-agnostic ; this confirms they don't trip
+on the existing macro-driven scroll FSM that owns the same struct.
+"""
+from __future__ import annotations
+
+import pytest
+
+from kintsuki import Button
+
+from _ff4kintsuki import load_emu_from_kss, kss_path, tap
+
+
+FIELD_MENU_ROLLING_BASE = 0x7E1BA8
+TREASURE_ROLLING_BASE = 0x7E1BD0
+
+# RollingBufferState offsets (mirrors src/items.i).
+DIRTY_MASK_OFFSET = 25
+
+
+@pytest.fixture
+def field_menu():
+    """Field menu Items submenu open, cursor on Aiguille d'or."""
+    e = load_emu_from_kss(kss_path("ff4-before-inventory-opens.kss"),
+                          settle_frames=60)
+    tap(e, Button.A)
+    e.run_frames(60)
+    yield e
+    e.close()
+
+
+def test_invalidate_all_lights_live_field_menu_state(field_menu) -> None:
+    """`rolling_engine_invalidate_all` on the live field-menu state base
+    sets dirty_mask = $FF without disturbing the rest of the struct."""
+    e = field_menu
+    # Snapshot the engine-scratch bytes (top_row..hdma_copy_pending) so
+    # we can confirm the engine entry leaves them alone.
+    pre = bytes(e.read(FIELD_MENU_ROLLING_BASE + i) for i in range(15))
+
+    fn = e.lookup_symbol_addr("rolling_engine.rolling_engine_invalidate_all")
+    assert fn is not None
+    e.call(fn, x=FIELD_MENU_ROLLING_BASE & 0xFFFF)
+
+    dirty = e.read(FIELD_MENU_ROLLING_BASE + DIRTY_MASK_OFFSET)
+    assert dirty == 0xFF, f"dirty_mask expected $FF, got ${dirty:02X}"
+
+    post = bytes(e.read(FIELD_MENU_ROLLING_BASE + i) for i in range(15))
+    assert pre == post, (
+        "engine-scratch bytes were touched by invalidate_all "
+        f"(pre={pre.hex()} post={post.hex()})")
+
+
+def test_vblank_flush_clears_live_dirty_mask(field_menu) -> None:
+    """Plant dirty_mask = $FF directly (no preceding engine call so the
+    field-menu NMI hooks can't intercept), then call vblank_flush and
+    verify the mask is cleared on live state."""
+    e = field_menu
+    e.write(FIELD_MENU_ROLLING_BASE + DIRTY_MASK_OFFSET, 0xFF)
+    assert e.read(FIELD_MENU_ROLLING_BASE + DIRTY_MASK_OFFSET) == 0xFF
+
+    flush = e.lookup_symbol_addr("rolling_engine.rolling_engine_vblank_flush")
+    e.call(flush, x=FIELD_MENU_ROLLING_BASE & 0xFFFF)
+    assert e.read(FIELD_MENU_ROLLING_BASE + DIRTY_MASK_OFFSET) == 0x00
+
+
+def test_field_menu_init_populates_engine_config(field_menu) -> None:
+    """init_menu_rolling_buffer_impl writes the phase-2 engine config +
+    hook far-ptrs into the field-menu state struct.
+
+    Confirms the per-menu init now arms the bank-20 engine entries with
+    everything they need (visible_rows, item_list_ptr, hdma_channel,
+    vwf_cfg_ptr, render/hdma/draw hooks) even though the macro is still
+    driving the actual scroll machinery.
+    """
+    e = field_menu
+    base = FIELD_MENU_ROLLING_BASE
+    assert e.read(base + 15) == 0x0A           # visible_rows = 10
+    assert e.read(base + 16) == 0x02           # slot_height_tiles = 2
+    assert e.read(base + 17) == 0x40           # item_list_ptr low
+    assert e.read(base + 18) == 0x14           # item_list_ptr mid
+    assert e.read(base + 19) == 0x7E           # item_list_ptr bank
+    assert e.read(base + 20) == 0x30           # item_count = 48
+    assert e.read(base + 21) == 0x05           # hdma_channel = 5
+    assert e.read(base + 22) == 0x80           # vwf_cfg_ptr low
+    assert e.read(base + 23) == 0x70           # vwf_cfg_ptr mid
+    assert e.read(base + 24) == 0x70           # vwf_cfg_ptr bank
+    # Hook far-ptrs land in bank-20 reloc region.
+    for hook_off in (26, 29, 32):
+        bank = e.read(base + hook_off + 2)
+        assert bank == 0x20, (
+            f"hook at +{hook_off} bank=${bank:02X}, expected $20")
+
+
+def test_cursor_down_on_field_menu_advances_slot_index(field_menu) -> None:
+    """`rolling_engine_cursor_down` advances slot_index on live state.
+
+    Plants visible_rows = 11 (MENU_BUFFER_SLOTS), slot_index = 3 ; calls
+    cursor_down ; expects 4. visible_rows + slot_index live in the
+    engine-extended portion of the struct so we can poke them
+    independently of the live engine state."""
+    e = field_menu
+    e.write(FIELD_MENU_ROLLING_BASE + 3, 3)   # slot_index
+    e.write(FIELD_MENU_ROLLING_BASE + 15, 11) # visible_rows
+
+    fn = e.lookup_symbol_addr("rolling_engine.rolling_engine_cursor_down")
+    e.call(fn, x=FIELD_MENU_ROLLING_BASE & 0xFFFF)
+
+    assert e.read(FIELD_MENU_ROLLING_BASE + 3) == 4
diff --git a/tests/test_rolling_engine_skeleton.py b/tests/test_rolling_engine_skeleton.py
new file mode 100644
index 0000000..6e3fbcd
--- /dev/null
+++ b/tests/test_rolling_engine_skeleton.py
@@ -0,0 +1,388 @@
+"""Phase-1 smoke for the unified rolling-inventory engine.
+
+Each entry currently RTLs immediately (plans/rolling_inventory_engine.md
+phase 1 stub) so the harness only verifies:
+
+* All eight bank-20 symbols resolve via the .adbg debug info.
+* Each entry, when JSL'd from a fresh emulator with X = a sentinel
+  state-ptr, returns cleanly without crashing or trashing X/Y.
+* `RollingBufferState` struct accepts the phase-1 extended fields
+  (visible_rows, slot_height_tiles, item_list_ptr, item_count,
+  hdma_channel, vwf_cfg_ptr, dirty_mask) without breaking the
+  existing field-items / treasure rolling consumers.
+
+Phase 2+ replaces the stubs with real implementations ; the same
+harness expands to drive a fake inventory and assert scroll-state
+transitions, dirty-mask flush ordering, and HDMA channel arming.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from _ff4kintsuki import load_emu_from_kss, kss_path
+
+# RollingBufferState field offsets (mirrors src/items.i `.struct
+# RollingBufferState`). Keep in sync if the struct evolves.
+RBS_SLOT_INDEX = 3
+RBS_VISIBLE_ROWS = 15
+RBS_DIRTY_MASK = 25
+
+# WRAM scratch region for engine state during unit tests. Sits in the
+# vanilla "$AAAA filler" area at $7E:0500 -- never touched by the rest
+# of the field-menu / treasure code paths during the menu settle so
+# reads + writes stay deterministic.
+SCRATCH_STATE = 0x7E0500
+
+
+ENGINE_ENTRIES = [
+    "rolling_engine.rolling_engine_init",
+    "rolling_engine.rolling_engine_tick",
+    "rolling_engine.rolling_engine_vblank_flush",
+    "rolling_engine.rolling_engine_cursor_up",
+    "rolling_engine.rolling_engine_cursor_down",
+    "rolling_engine.rolling_engine_invalidate_slot",
+    "rolling_engine.rolling_engine_invalidate_all",
+    "rolling_engine.rolling_engine_shutdown",
+]
+
+
+@pytest.fixture
+def emu():
+    e = load_emu_from_kss(kss_path("ff4-before-inventory-opens.kss"),
+                          settle_frames=10)
+    yield e
+    e.close()
+
+
+def test_engine_entries_resolve(emu) -> None:
+    """Every phase-1 entry point is exported via the .adbg file."""
+    missing = [sym for sym in ENGINE_ENTRIES
+               if emu.lookup_symbol_addr(sym) is None]
+    assert not missing, f"missing engine symbols: {missing}"
+
+
+def test_engine_entries_in_bank_20(emu) -> None:
+    """Phase-1 entries land in the bank-20 reloc region ($20:8000+)."""
+    for sym in ENGINE_ENTRIES:
+        addr = emu.lookup_symbol_addr(sym)
+        assert addr is not None, sym
+        assert (addr >> 16) == 0x20, (
+            f"{sym} at ${addr:06X} should be in bank 20")
+
+
+# --------------------------------------------------------------- Helpers
+
+
+def _setup_state(emu, *, slot_index: int, visible_rows: int,
+                 dirty_mask: int = 0) -> int:
+    """Plant a sentinel `RollingBufferState` at SCRATCH_STATE and return
+    the low-16 of its address (= the value that goes in X for engine
+    JSLs ; the bank-$7E implied addressing in the engine reads from
+    $7E:<low16> + struct-field offset).
+    """
+    emu.write(SCRATCH_STATE + RBS_SLOT_INDEX, slot_index)
+    emu.write(SCRATCH_STATE + RBS_VISIBLE_ROWS, visible_rows)
+    emu.write(SCRATCH_STATE + RBS_DIRTY_MASK, dirty_mask)
+    return SCRATCH_STATE & 0xFFFF
+
+
+def _read_slot_index(emu) -> int:
+    return emu.read(SCRATCH_STATE + RBS_SLOT_INDEX)
+
+
+def _read_dirty_mask(emu) -> int:
+    return emu.read(SCRATCH_STATE + RBS_DIRTY_MASK)
+
+
+# --------------------------------------------------------------- Cursor
+
+
+def test_cursor_down_increments_slot_index(emu) -> None:
+    """cursor_down advances slot_index by 1 while inside the window."""
+    x = _setup_state(emu, slot_index=2, visible_rows=10)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_cursor_down")
+    emu.call(fn, x=x)
+    assert _read_slot_index(emu) == 3
+
+
+def test_cursor_down_wraps_to_zero(emu) -> None:
+    """cursor_down on the last slot wraps back to 0."""
+    x = _setup_state(emu, slot_index=9, visible_rows=10)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_cursor_down")
+    emu.call(fn, x=x)
+    assert _read_slot_index(emu) == 0
+
+
+def test_cursor_up_decrements_slot_index(emu) -> None:
+    """cursor_up advances slot_index backwards by 1."""
+    x = _setup_state(emu, slot_index=4, visible_rows=10)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_cursor_up")
+    emu.call(fn, x=x)
+    assert _read_slot_index(emu) == 3
+
+
+def test_cursor_up_wraps_from_zero(emu) -> None:
+    """cursor_up at slot 0 wraps to (visible_rows - 1)."""
+    x = _setup_state(emu, slot_index=0, visible_rows=10)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_cursor_up")
+    emu.call(fn, x=x)
+    assert _read_slot_index(emu) == 9
+
+
+# --------------------------------------------------------------- Dirty
+
+
+def test_invalidate_slot_sets_bit(emu) -> None:
+    """invalidate_slot sets bit (1 << A) in dirty_mask without disturbing
+    bits already lit."""
+    x = _setup_state(emu, slot_index=0, visible_rows=10, dirty_mask=0x01)
+    fn = emu.lookup_symbol_addr(
+        "rolling_engine.rolling_engine_invalidate_slot")
+    emu.call(fn, a=4, x=x)
+    # bit 4 set, bit 0 from setup also set
+    assert _read_dirty_mask(emu) == 0x11
+
+
+def test_invalidate_slot_idempotent(emu) -> None:
+    """Setting an already-dirty bit leaves dirty_mask unchanged."""
+    x = _setup_state(emu, slot_index=0, visible_rows=10, dirty_mask=0x20)
+    fn = emu.lookup_symbol_addr(
+        "rolling_engine.rolling_engine_invalidate_slot")
+    emu.call(fn, a=5, x=x)
+    assert _read_dirty_mask(emu) == 0x20
+
+
+def test_invalidate_all_lights_full_mask(emu) -> None:
+    """invalidate_all sets dirty_mask to $FF."""
+    x = _setup_state(emu, slot_index=0, visible_rows=10, dirty_mask=0x00)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_invalidate_all")
+    emu.call(fn, x=x)
+    assert _read_dirty_mask(emu) == 0xFF
+
+
+# --------------------------------------------------------------- Init
+
+
+# Engine-scratch bytes that init must zero (offsets within state struct).
+INIT_ZERO_BYTES = [
+    0,   # top_row
+    1,   # buffer_pos
+    2,   # edge_row
+    3,   # slot_index
+    6,   # hdma_enable
+    7,   # _pad
+    8,   # scroll_state
+    9,   # scroll_remaining
+    10,  # scroll_direction
+    11,  # transfer_pending
+    12,  # scroll_anim_offset lo
+    13,  # scroll_anim_offset hi
+    14,  # hdma_copy_pending
+    25,  # dirty_mask
+]
+
+
+def _fill_state_with_pattern(emu, byte: int = 0xAA, length: int = 32) -> None:
+    """Pre-fill 32 bytes at SCRATCH_STATE with a known sentinel so init's
+    zero-out and base_scroll = $FFFF stamping is visible against the
+    pattern."""
+    for i in range(length):
+        emu.write(SCRATCH_STATE + i, byte)
+
+
+def test_init_zeroes_engine_scratch(emu) -> None:
+    """rolling_engine_init clears the 14 engine-scratch + dirty_mask bytes."""
+    _fill_state_with_pattern(emu)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_init")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+    for off in INIT_ZERO_BYTES:
+        val = emu.read(SCRATCH_STATE + off)
+        assert val == 0x00, (
+            f"offset {off} expected zero after init, got ${val:02X}")
+
+
+def test_init_stamps_base_scroll_sentinel(emu) -> None:
+    """rolling_engine_init writes $FFFF into base_scroll (offsets 4 + 5)."""
+    _fill_state_with_pattern(emu)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_init")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+    assert emu.read(SCRATCH_STATE + 4) == 0xFF
+    assert emu.read(SCRATCH_STATE + 5) == 0xFF
+
+
+def test_init_preserves_config_fields(emu) -> None:
+    """Caller-managed config fields (visible_rows + dirty_mask aside)
+    must NOT be touched by init : engine_init clears scratch only."""
+    _fill_state_with_pattern(emu)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_init")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+    # visible_rows (offset 15) + item_list_ptr 17..19 + item_count 20 +
+    # hdma_channel 21 + vwf_cfg_ptr 22..24 should keep the pattern.
+    for off in [15, 16, 17, 18, 19, 20, 21, 22, 23, 24]:
+        val = emu.read(SCRATCH_STATE + off)
+        assert val == 0xAA, (
+            f"config offset {off} got clobbered by init: ${val:02X}")
+
+
+# --------------------------------------------------------------- Shutdown
+
+
+def test_shutdown_clears_runtime_flags(emu) -> None:
+    """rolling_engine_shutdown clears hdma_enable / dirty_mask /
+    scroll_state / transfer_pending / hdma_copy_pending and restamps
+    the base_scroll sentinel."""
+    _fill_state_with_pattern(emu)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_shutdown")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+    assert emu.read(SCRATCH_STATE + 6) == 0x00   # hdma_enable
+    assert emu.read(SCRATCH_STATE + 8) == 0x00   # scroll_state
+    assert emu.read(SCRATCH_STATE + 11) == 0x00  # transfer_pending
+    assert emu.read(SCRATCH_STATE + 14) == 0x00  # hdma_copy_pending
+    assert emu.read(SCRATCH_STATE + 25) == 0x00  # dirty_mask
+    assert emu.read(SCRATCH_STATE + 4) == 0xFF   # base_scroll lo
+    assert emu.read(SCRATCH_STATE + 5) == 0xFF   # base_scroll hi
+
+
+def test_shutdown_leaves_cursor_alone(emu) -> None:
+    """Shutdown does not nuke slot_index / buffer_pos / edge_row /
+    top_row ; those are caller's UI state."""
+    _setup_state(emu, slot_index=4, visible_rows=10)
+    emu.write(SCRATCH_STATE + 0, 0x07)  # top_row
+    emu.write(SCRATCH_STATE + 1, 0x03)  # buffer_pos
+    emu.write(SCRATCH_STATE + 2, 0x02)  # edge_row
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_shutdown")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+    assert emu.read(SCRATCH_STATE + 0) == 0x07
+    assert emu.read(SCRATCH_STATE + 1) == 0x03
+    assert emu.read(SCRATCH_STATE + 2) == 0x02
+    assert emu.read(SCRATCH_STATE + 3) == 0x04  # slot_index
+
+
+# --------------------------------------------------------------- Tick
+
+
+# Scroll FSM byte offsets within RollingBufferState.
+RBS_SCROLL_STATE = 8
+RBS_SCROLL_REMAINING = 9
+RBS_TRANSFER_PENDING = 11
+
+
+def _setup_scroll(emu, *, state: int, remaining: int) -> int:
+    emu.write(SCRATCH_STATE + RBS_SCROLL_STATE, state)
+    emu.write(SCRATCH_STATE + RBS_SCROLL_REMAINING, remaining)
+    emu.write(SCRATCH_STATE + RBS_TRANSFER_PENDING, 0)
+    return SCRATCH_STATE & 0xFFFF
+
+
+def test_tick_idle_is_a_noop(emu) -> None:
+    """Tick with scroll_state == 0 leaves scroll_remaining untouched."""
+    x = _setup_scroll(emu, state=0, remaining=8)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_tick")
+    emu.call(fn, x=x)
+    assert emu.read(SCRATCH_STATE + RBS_SCROLL_STATE) == 0
+    assert emu.read(SCRATCH_STATE + RBS_SCROLL_REMAINING) == 8
+    assert emu.read(SCRATCH_STATE + RBS_TRANSFER_PENDING) == 0
+
+
+def test_tick_decrements_scroll_remaining(emu) -> None:
+    """Tick with scroll_state != 0 drops scroll_remaining by 1."""
+    x = _setup_scroll(emu, state=1, remaining=4)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_tick")
+    emu.call(fn, x=x)
+    assert emu.read(SCRATCH_STATE + RBS_SCROLL_REMAINING) == 3
+    assert emu.read(SCRATCH_STATE + RBS_SCROLL_STATE) == 1
+    assert emu.read(SCRATCH_STATE + RBS_TRANSFER_PENDING) == 0
+
+
+def test_tick_finishes_animation_on_last_frame(emu) -> None:
+    """When scroll_remaining hits zero, scroll_state clears and
+    transfer_pending stamps 1 so NMI flush picks up the final frame."""
+    x = _setup_scroll(emu, state=1, remaining=1)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_tick")
+    emu.call(fn, x=x)
+    assert emu.read(SCRATCH_STATE + RBS_SCROLL_REMAINING) == 0
+    assert emu.read(SCRATCH_STATE + RBS_SCROLL_STATE) == 0
+    assert emu.read(SCRATCH_STATE + RBS_TRANSFER_PENDING) == 1
+
+
+# --------------------------------------------------------------- Flush
+
+
+def test_vblank_flush_clears_dirty_mask(emu) -> None:
+    """vblank_flush clears state.dirty_mask after handling the dirty rows."""
+    x = _setup_state(emu, slot_index=0, visible_rows=10, dirty_mask=0x35)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_vblank_flush")
+    emu.call(fn, x=x)
+    assert _read_dirty_mask(emu) == 0x00
+
+
+def test_vblank_flush_noop_when_clean(emu) -> None:
+    """vblank_flush is harmless when dirty_mask is already zero."""
+    x = _setup_state(emu, slot_index=0, visible_rows=10, dirty_mask=0x00)
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_vblank_flush")
+    emu.call(fn, x=x)
+    assert _read_dirty_mask(emu) == 0x00
+
+
+def test_init_fires_fn_draw_window_when_armed(emu) -> None:
+    """When state.fn_draw_window is non-zero, rolling_engine_init JSLs
+    through it.
+
+    Plant a small RTL stub at $7E:0600 that writes a sentinel ($A5) to a
+    witness byte ($7E:0700), point fn_draw_window at the stub, then
+    call init. The witness byte should flip if the hook actually fires.
+    Stub bytes : `lda #$A5 ; sta.l $7E:0700 ; rtl` = A9 A5 8F 00 07 7E 6B.
+    """
+    stub_bytes = bytes([0xA9, 0xA5, 0x8F, 0x00, 0x07, 0x7E, 0x6B])
+    for i, b in enumerate(stub_bytes):
+        emu.write(0x7E0600 + i, b)
+    emu.write(0x7E0700, 0x00)
+
+    _fill_state_with_pattern(emu)
+    # fn_draw_window lives at struct offset 32..34 (after dirty_mask).
+    emu.write(SCRATCH_STATE + 32, 0x00)
+    emu.write(SCRATCH_STATE + 33, 0x06)
+    emu.write(SCRATCH_STATE + 34, 0x7E)
+
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_init")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+
+    witness = emu.read(0x7E0700)
+    assert witness == 0xA5, (
+        f"fn_draw_window stub did not fire ; witness=${witness:02X}")
+
+
+def test_init_loops_fn_render_slot_visible_rows_times(emu) -> None:
+    """rolling_engine_init calls fn_render_slot once per visible row,
+    planting slot_index = edge_row = iteration count before each call.
+
+    Stub at $7E:0800 increments a counter at $7E:0900 then RTLs.
+    With visible_rows = 5, init should fire 5 callbacks (counter = 5)
+    and leave slot_index + edge_row pinned at 4 (last iteration).
+    """
+    # Stub : lda.l $7E0900 ; inc ; sta.l $7E0900 ; rtl
+    #        AF 00 09 7E   ; 1A   ; 8F 00 09 7E   ; 6B
+    stub = bytes([0xAF, 0x00, 0x09, 0x7E,
+                  0x1A,
+                  0x8F, 0x00, 0x09, 0x7E,
+                  0x6B])
+    for i, b in enumerate(stub):
+        emu.write(0x7E0800 + i, b)
+    emu.write(0x7E0900, 0x00)
+
+    # Zero the whole scratch state then set just visible_rows + render hook.
+    for i in range(40):
+        emu.write(SCRATCH_STATE + i, 0x00)
+    emu.write(SCRATCH_STATE + RBS_VISIBLE_ROWS, 5)
+    emu.write(SCRATCH_STATE + 26, 0x00)  # fn_render_slot low
+    emu.write(SCRATCH_STATE + 27, 0x08)  # fn_render_slot mid
+    emu.write(SCRATCH_STATE + 28, 0x7E)  # fn_render_slot bank
+
+    fn = emu.lookup_symbol_addr("rolling_engine.rolling_engine_init")
+    emu.call(fn, x=SCRATCH_STATE & 0xFFFF)
+
+    counter = emu.read(0x7E0900)
+    assert counter == 5, f"fn_render_slot fired {counter} times, expected 5"
+    assert emu.read(SCRATCH_STATE + 3) == 4   # slot_index
+    assert emu.read(SCRATCH_STATE + 2) == 4   # edge_row
diff --git a/tests/test_vwf_asset_widths.py b/tests/test_vwf_asset_widths.py
new file mode 100644
index 0000000..7cf281a
--- /dev/null
+++ b/tests/test_vwf_asset_widths.py
@@ -0,0 +1,67 @@
+"""VWF asset width budget tests.
+
+For each fixed-list XML asset that gets rendered through the 8x8 menu
+VWF blitter, measure every entry's rendered tile-width with the same
+font + kerning tables the runtime uses, and assert it fits the slot
+budget the rendering code reserves.
+
+Battle inventory uses 10 VWF tile_ids per slot (the symbol byte +
+colon + 2 digits are fixed-width tiles outside the budget). The
+allocator clamp freezes at slot_base+9 ; an 11+ tile name would
+truncate at render time.
+"""
+from __future__ import annotations
+import math
+import sys
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+
+from script import Table  # noqa: E402
+from metrics import TextMetrics  # noqa: E402
+from build import read_fixed_from_xml  # noqa: E402
+
+
+def _load(asset_xml: Path) -> tuple[list, TextMetrics, Table]:
+    table = Table(str(ROOT / "text" / "ff4_menus.tbl"))
+    metrics = TextMetrics(table, [str(ROOT / "assets" / "menu_font.dat")], char_height=8)
+    items = read_fixed_from_xml(
+        str(asset_xml),
+        table,
+        formatter=lambda t: (t or "").strip() + "[end]",
+    )
+    return items, metrics, table
+
+
+@pytest.mark.parametrize(
+    "asset_xml, vwf_tile_budget, kind",
+    [
+        ("text/fr/items.xml",            10, "battle inventory items"),
+        ("text/fr/items_unleashed.xml",  10, "battle inventory items_unleashed"),
+    ],
+)
+def test_vwf_asset_within_budget(asset_xml: str, vwf_tile_budget: int, kind: str) -> None:
+    items, metrics, table = _load(ROOT / asset_xml)
+    offenders = []
+    for i, ptr in enumerate(items):
+        body = ptr.value
+        # First byte is the fixed-width symbol icon ; VWF starts at byte 1.
+        vwf_bytes = body[1:] if body else b""
+        # Strip trailing [end] sentinel before measurement.
+        if vwf_bytes and vwf_bytes[-1] == 0:
+            vwf_bytes = vwf_bytes[:-1]
+        width_px = metrics.measure_bytes(vwf_bytes)
+        tiles = math.ceil(width_px / 8)
+        if tiles > vwf_tile_budget:
+            text = table.to_text(body).replace("[end]", "")
+            offenders.append((i, tiles, width_px, text))
+    assert not offenders, (
+        f"{kind}: {len(offenders)} entries exceed {vwf_tile_budget}-tile VWF budget:\n  "
+        + "\n  ".join(
+            f"id={i:>3} {tiles:>2} tiles {px:>3}px  {text!r}"
+            for i, tiles, px, text in offenders
+        )
+    )
diff --git a/text/fr/items_unleashed.xml b/text/fr/items_unleashed.xml
new file mode 100644
index 0000000..809d9d1
--- /dev/null
+++ b/text/fr/items_unleashed.xml
@@ -0,0 +1,259 @@
+<?xml version="1.0" encoding="utf-8" standalone="yes"?>
+<sn:fixedlist xmlns:sn="http://snes.ninja/ScriptNS" length="17" padding=" ">
+    <sn:fixed></sn:fixed>
+    <sn:fixed>[0x29]Griffe de feu</sn:fixed>
+    <sn:fixed>[0x29]Griffe de glace</sn:fixed>
+    <sn:fixed>[0x29]Griffe de foudre</sn:fixed>
+    <sn:fixed>[0x29]Griffe de fée</sn:fixed>
+    <sn:fixed>[0x29]Griffe d'enfer</sn:fixed>
+    <sn:fixed>[0x29]Griffe de chat</sn:fixed>
+    <sn:fixed>[0x2a]Baguette</sn:fixed>
+    <sn:fixed>[0x2a]Baguette glace</sn:fixed>
+    <sn:fixed>[0x2a]Baguette feu</sn:fixed>
+    <sn:fixed>[0x2a]Baguette foudre</sn:fixed>
+    <sn:fixed>[0x2a]Baguette de mue</sn:fixed>
+    <sn:fixed>[0x2a]Baguette de fée</sn:fixed>
+    <sn:fixed>[0x2a]Baguette étoiles</sn:fixed>
+    <sn:fixed>[0x2a]Baguette Lilith</sn:fixed>
+    <sn:fixed>[0x2b]Bâton</sn:fixed>
+    <sn:fixed>[0x2b]Bâton de soin</sn:fixed>
+    <sn:fixed>[0x2b]Bâton mithril</sn:fixed>
+    <sn:fixed>[0x2b]Bâton de force</sn:fixed>
+    <sn:fixed>[0x2b]Bâton pulsar</sn:fixed>
+    <sn:fixed>[0x2b]Bâton du sage</sn:fixed>
+    <sn:fixed>[0x2b]Bâton runique</sn:fixed>
+    <sn:fixed>[0x2c]Epée ténébreuse</sn:fixed>
+    <sn:fixed>[0x2c]Lame de l'ombre</sn:fixed>
+    <sn:fixed>[0x2c]Death Bringer</sn:fixed>
+    <sn:fixed>[0x2e]Epée légendaire</sn:fixed>
+    <sn:fixed>[0x2e]Epée de lumière</sn:fixed>
+    <sn:fixed>[0x2e]Excalibur</sn:fixed>
+    <sn:fixed>[0x2d]Epée de flammes</sn:fixed>
+    <sn:fixed>[0x2d]Lame de glace</sn:fixed>
+    <sn:fixed>[0x2d]Défenseur</sn:fixed>
+    <sn:fixed>[0x2d]Epée de sang</sn:fixed>
+    <sn:fixed>[0x2d]Epée ancienne</sn:fixed>
+    <sn:fixed>[0x2d]Epée du sommeil</sn:fixed>
+    <sn:fixed>[0x2d]Lame briseuse</sn:fixed>
+    <sn:fixed>[0x2f]Lance</sn:fixed>
+    <sn:fixed>[0x2f]Lance du vent</sn:fixed>
+    <sn:fixed>[0x2f]Lance de feu</sn:fixed>
+    <sn:fixed>[0x2f]Lance de glace</sn:fixed>
+    <sn:fixed>[0x2f]Lance du dragon</sn:fixed>
+    <sn:fixed>[0x2f]Lance sacrée</sn:fixed>
+    <sn:fixed>[0x2f]Lance de sang</sn:fixed>
+    <sn:fixed>[0x2f]Gungnir</sn:fixed>
+    <sn:fixed>[0x31]Kunai</sn:fixed>
+    <sn:fixed>[0x31]Ashura</sn:fixed>
+    <sn:fixed>[0x31]Kotetsu</sn:fixed>
+    <sn:fixed>[0x31]Kikuichimonji</sn:fixed>
+    <sn:fixed>[0x31]Murasame</sn:fixed>
+    <sn:fixed>[0x31]Masamune</sn:fixed>
+    <sn:fixed>[0x30]Dague Assassin</sn:fixed>
+    <sn:fixed>[0x30]Tueur de mages</sn:fixed>
+    <sn:fixed>[0x3a]Fouet</sn:fixed>
+    <sn:fixed>[0x3a]Fouet à chaîne</sn:fixed>
+    <sn:fixed>[0x3a]Fouet foudre</sn:fixed>
+    <sn:fixed>[0x3a]Fouet de feu</sn:fixed>
+    <sn:fixed>[0x3a]Crin de dragon</sn:fixed>
+    <sn:fixed>[0x34]Hachette</sn:fixed>
+    <sn:fixed>[0x34]Hache des nains</sn:fixed>
+    <sn:fixed>[0x34]Tueur d'ogres</sn:fixed>
+    <sn:fixed>[0x30]Dague mithril</sn:fixed>
+    <sn:fixed>[0x30]Dague dansante</sn:fixed>
+    <sn:fixed>[0x2d]Epée mithril</sn:fixed>
+    <sn:fixed>[0x30]Couperet</sn:fixed>
+    <sn:fixed>[0x2e]Ragnarok</sn:fixed>
+    <sn:fixed>[0x32]Shuriken</sn:fixed>
+    <sn:fixed>[0x32]Shuriken Fuma</sn:fixed>
+    <sn:fixed>[0x33]Boomerang</sn:fixed>
+    <sn:fixed>[0x33]Engetsurin</sn:fixed>
+    <sn:fixed>[0x36]Harpe des rêves</sn:fixed>
+    <sn:fixed>[0x36]Harpe de Lamia</sn:fixed>
+    <sn:fixed></sn:fixed>
+    <sn:fixed>[0x34]Hache d'Idun</sn:fixed>
+    <sn:fixed>[0x34]Hache runique</sn:fixed>
+    <sn:fixed>[0x35]Marteau mithril</sn:fixed>
+    <sn:fixed>[0x35]Marteau Terre</sn:fixed>
+    <sn:fixed>[0x35]Maillet</sn:fixed>
+    <sn:fixed>[0x2d]Vengeur</sn:fixed>
+    <sn:fixed>[0x37]Arc</sn:fixed>
+    <sn:fixed>[0x37]Arbalète</sn:fixed>
+    <sn:fixed>[0x37]Grand arc</sn:fixed>
+    <sn:fixed>[0x37]Arc mortel</sn:fixed>
+    <sn:fixed>[0x37]Arc elfique</sn:fixed>
+    <sn:fixed>[0x37]Arc de Yoichi</sn:fixed>
+    <sn:fixed>[0x37]Arc d'Artémis</sn:fixed>
+    <sn:fixed>[0x38]Flèche de fer</sn:fixed>
+    <sn:fixed>[0x38]Flèche sacrée</sn:fixed>
+    <sn:fixed>[0x38]Flèche de feu</sn:fixed>
+    <sn:fixed>[0x38]Flèche de glace</sn:fixed>
+    <sn:fixed>[0x38]Flèche foudre</sn:fixed>
+    <sn:fixed>[0x38]Flèche ténèbres</sn:fixed>
+    <sn:fixed>[0x38]Flèche de poison</sn:fixed>
+    <sn:fixed>[0x38]Flèche muselière</sn:fixed>
+    <sn:fixed>[0x38]Flèche d'ange</sn:fixed>
+    <sn:fixed>[0x38]Flèche Yoichi</sn:fixed>
+    <sn:fixed>[0x38]Flèche Méduse</sn:fixed>
+    <sn:fixed>[0x38]Flèche Artémis</sn:fixed>
+    <sn:fixed></sn:fixed>
+    <sn:fixed>[0x3b]Bouclier de fer</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier noir</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier démon</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier sacré</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier mithril</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier flamme</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier glace</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier diamant</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier d'Egide</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier Genji</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier dragon</sn:fixed>
+    <sn:fixed>[0x3b]Bouclier cristal</sn:fixed>
+    <sn:fixed>[0x3c]Heaume de fer</sn:fixed>
+    <sn:fixed>[0x3c]Heaume noir</sn:fixed>
+    <sn:fixed>[0x3c]Heaume Hadès</sn:fixed>
+    <sn:fixed>[0x3c]Heaume démon</sn:fixed>
+    <sn:fixed>[0x3c]Heaume sacré</sn:fixed>
+    <sn:fixed>[0x3c]Heaume mithril</sn:fixed>
+    <sn:fixed>[0x3c]Heaume diamant</sn:fixed>
+    <sn:fixed>[0x3c]Heaume Genji</sn:fixed>
+    <sn:fixed>[0x3c]Heaume dragon</sn:fixed>
+    <sn:fixed>[0x3c]Heaume cristal</sn:fixed>
+    <sn:fixed>[0x3c]Chapeau de cuir</sn:fixed>
+    <sn:fixed>[0x3c]Chapeau plumes</sn:fixed>
+    <sn:fixed>[0x3c]Tricorne</sn:fixed>
+    <sn:fixed>[0x3c]Mitre du prêtre</sn:fixed>
+    <sn:fixed>[0x3c]Bandeau d'or</sn:fixed>
+    <sn:fixed>[0x3c]Ruban</sn:fixed>
+    <sn:fixed>[0x3c]Bandana</sn:fixed>
+    <sn:fixed>[0x3c]Béret vert</sn:fixed>
+    <sn:fixed>[0x3c]Capuche noire</sn:fixed>
+    <sn:fixed>[0x3c]Masque de verre</sn:fixed>
+    <sn:fixed>[0x3d]Armure de fer</sn:fixed>
+    <sn:fixed>[0x3d]Armure noire</sn:fixed>
+    <sn:fixed>[0x3d]Armure Hadès</sn:fixed>
+    <sn:fixed>[0x3d]Armure démon</sn:fixed>
+    <sn:fixed>[0x3d]Armure Chevalier</sn:fixed>
+    <sn:fixed>[0x3d]Armure mithril</sn:fixed>
+    <sn:fixed>[0x3d]Cotte de flammes</sn:fixed>
+    <sn:fixed>[0x3d]Armure de glace</sn:fixed>
+    <sn:fixed>[0x3d]Armure diamant</sn:fixed>
+    <sn:fixed>[0x3d]Armure Genji</sn:fixed>
+    <sn:fixed>[0x3d]Cotte de dragon</sn:fixed>
+    <sn:fixed>[0x3d]Cotte de cristal</sn:fixed>
+    <sn:fixed>[0x3d]Vêtements</sn:fixed>
+    <sn:fixed>[0x3d]Tunique cuir</sn:fixed>
+    <sn:fixed>[0x3d]Robe de Gaïa</sn:fixed>
+    <sn:fixed>[0x3d]Robe du prêtre</sn:fixed>
+    <sn:fixed>[0x3d]Robe noire</sn:fixed>
+    <sn:fixed>[0x3d]Robe de lumière</sn:fixed>
+    <sn:fixed>[0x3d]Robe blanche</sn:fixed>
+    <sn:fixed>[0x3d]Ruban de force</sn:fixed>
+    <sn:fixed>[0x3d]Bustier Minerva</sn:fixed>
+    <sn:fixed>[0x3d]Habit bagnard</sn:fixed>
+    <sn:fixed>[0x3d]Habit poète</sn:fixed>
+    <sn:fixed>[0x3d]Tenue Kenpou</sn:fixed>
+    <sn:fixed>[0x3d]Ceinture noire</sn:fixed>
+    <sn:fixed>[0x3d]Armure adamant</sn:fixed>
+    <sn:fixed>[0x3d]Habit ninja</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet fer</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet noir</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet Hadès</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet démon</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet mithril</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet diamant</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet géant</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet Genji</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet dragon</sn:fixed>
+    <sn:fixed>[0x3e]Gantelet cristal</sn:fixed>
+    <sn:fixed>[0x3e]Brassard de fer</sn:fixed>
+    <sn:fixed>[0x3e]Anneau rubis</sn:fixed>
+    <sn:fixed>[0x3e]Brassard argent</sn:fixed>
+    <sn:fixed>[0x3e]Brassard force</sn:fixed>
+    <sn:fixed>[0x3e]Brassard rune</sn:fixed>
+    <sn:fixed>[0x3e]Anneau cristal</sn:fixed>
+    <sn:fixed>[0x3e]Brassard diamant</sn:fixed>
+    <sn:fixed>[0x3e]Anneau garde</sn:fixed>
+    <sn:fixed>[0x3e]Anneau maudit</sn:fixed>
+    <sn:fixed>[0xff]Eclat de Bombe</sn:fixed>
+    <sn:fixed>[0xff]Bras de Bombe</sn:fixed>
+    <sn:fixed>[0xff]Vent du Sud</sn:fixed>
+    <sn:fixed>[0xff]Vent du Nord</sn:fixed>
+    <sn:fixed>[0xff]Foudre de Zeus</sn:fixed>
+    <sn:fixed>[0xff]Foudre divine</sn:fixed>
+    <sn:fixed>[0xff]Poussière étoile</sn:fixed>
+    <sn:fixed>[0xff]Baiser Lilith</sn:fixed>
+    <sn:fixed>[0xff]Croc vampire</sn:fixed>
+    <sn:fixed>[0xff]Vin de Bacchus</sn:fixed>
+    <sn:fixed>[0xff]Sandales Hermès</sn:fixed>
+    <sn:fixed>[0xff]Sablier cuivre</sn:fixed>
+    <sn:fixed>[0xff]Sablier argent</sn:fixed>
+    <sn:fixed>[0xff]Sablier d'or</sn:fixed>
+    <sn:fixed>[0xff]Toile d'araignée</sn:fixed>
+    <sn:fixed>[0xff]Bouc émissaire</sn:fixed>
+    <sn:fixed>[0xff]Croc rouge</sn:fixed>
+    <sn:fixed>[0xff]Croc blanc</sn:fixed>
+    <sn:fixed>[0xff]Croc bleu</sn:fixed>
+    <sn:fixed>[0xff]Rideau lumière</sn:fixed>
+    <sn:fixed>[0xff]Âme de Bombe</sn:fixed>
+    <sn:fixed>[0xff]Rideau lunaire</sn:fixed>
+    <sn:fixed>[0xff]Cloche silence</sn:fixed>
+    <sn:fixed>[0xff]Tambour terre</sn:fixed>
+    <sn:fixed>[0xff]Cristal</sn:fixed>
+    <sn:fixed>[0xff]Moustache Coeurl</sn:fixed>
+    <sn:fixed>[0xff]Grimoire</sn:fixed>
+    <sn:fixed>[0xff]Bestiaire</sn:fixed>
+    <sn:fixed>[0xff]Réveil</sn:fixed>
+    <sn:fixed>[0xff]Corne licorne</sn:fixed>
+    <sn:fixed>[0xff]Potion</sn:fixed>
+    <sn:fixed>[0xff]Hi-Potion</sn:fixed>
+    <sn:fixed>[0xff]Méga-Potion</sn:fixed>
+    <sn:fixed>[0xff]Ether</sn:fixed>
+    <sn:fixed>[0xff]Ether sec</sn:fixed>
+    <sn:fixed>[0xff]Elixir</sn:fixed>
+    <sn:fixed>[0xff]Plume de phénix</sn:fixed>
+    <sn:fixed>[0xff]Aiguille d'or</sn:fixed>
+    <sn:fixed>[0xff]Baiser de fée</sn:fixed>
+    <sn:fixed>[0xff]Marteau magique</sn:fixed>
+    <sn:fixed>[0xff]Aliment diète</sn:fixed>
+    <sn:fixed>[0xff]Herbe d'écho</sn:fixed>
+    <sn:fixed>[0xff]Collyre</sn:fixed>
+    <sn:fixed>[0xff]Antidote</sn:fixed>
+    <sn:fixed>[0xff]Crucifix</sn:fixed>
+    <sn:fixed>[0xff]Panacée</sn:fixed>
+    <sn:fixed>[0xff]Alarme</sn:fixed>
+    <sn:fixed>[0xff]Pomme d'or</sn:fixed>
+    <sn:fixed>[0xff]Pomme d'argent</sn:fixed>
+    <sn:fixed>[0xff]Goutte de Soma</sn:fixed>
+    <sn:fixed>[0xff]Tente</sn:fixed>
+    <sn:fixed>[0xff]Chalet</sn:fixed>
+    <sn:fixed>[0xff]Revue coquine</sn:fixed>
+    <sn:fixed>[0xff]Porte de secours</sn:fixed>
+    <sn:fixed>[0xff]Pain de nain</sn:fixed>
+    <sn:fixed>[0x41]Gobelin</sn:fixed>
+    <sn:fixed>[0x41]Bombe</sn:fixed>
+    <sn:fixed>[0x41]Cocatrix</sn:fixed>
+    <sn:fixed>[0x41]Flagelleur</sn:fixed>
+    <sn:fixed>[0xff]Légume Gysahl</sn:fixed>
+    <sn:fixed>[0xff]Laissez-passer</sn:fixed>
+    <sn:fixed>[0xff]Flûte Gysahl</sn:fixed>
+    <sn:fixed>[0xff]Anneau Bombe</sn:fixed>
+    <sn:fixed>[0xff]Clé de Baron</sn:fixed>
+    <sn:fixed>[0xff]Lumière désert</sn:fixed>
+    <sn:fixed>[0xff]Cristal terre</sn:fixed>
+    <sn:fixed>[0xff]Pierre de magma</sn:fixed>
+    <sn:fixed>[0xff]Collier de Luca</sn:fixed>
+    <sn:fixed>[0xff]Murmure</sn:fixed>
+    <sn:fixed>[0xff]Cristal ténèbres</sn:fixed>
+    <sn:fixed>[0xff]Queue de rat</sn:fixed>
+    <sn:fixed>[0xff]Adamantite</sn:fixed>
+    <sn:fixed>[0xff]Poêle d'amour</sn:fixed>
+    <sn:fixed>[0xff]Queue rose</sn:fixed>
+    <sn:fixed>[0xff]Clé de Lugae</sn:fixed>
+    <sn:fixed>[0xff]Matière sombre</sn:fixed>
+    <sn:fixed>[0xff]Scénario 0015</sn:fixed>
+    <sn:fixed>[0xff]Scénario 0016</sn:fixed>
+    <sn:fixed>[0xff]--Trier--</sn:fixed>
+    <sn:fixed>[0xff]Poubelle</sn:fixed>
+</sn:fixedlist>
diff --git a/todo/nmi-side-anim.md b/todo/nmi-side-anim.md
index c78cf75..df73ae9 100644
--- a/todo/nmi-side-anim.md
+++ b/todo/nmi-side-anim.md
@@ -99,7 +99,7 @@ update the first 8 bytes (HDMA tables often repeat).
 
 - `tests/_profile/profile_bank20.py` : measure dma_transfer
   inclusive cycle delta after each phase
-- Visual smoke on `ff4-battle-ext.kss` : Zu wing flap, char float
+- Visual smoke on `ff4-battle.kss` : Zu wing flap, char float
   state if any party member has Float
 
 ## Crack / risk