TH08 fidelity: skip et_ex on spawn-end tick; Stage-1 hit 3205→3214 by AgentMystia · Pull Request #1 · AgentMystia/th08_web

AgentMystia · 2026-08-24T13:14:57Z

Summary

Engine-level Stage 1+2 / Border Team fidelity against Th08.exe v1.00d. No phase clamps, hitbox shrinks, RNG special-cases, or per-frame workarounds. Wine/ptrace stays closed.

This pass (`0c395d6`)

Spawn-state ending ticks skip et_ex. TH08 OnUpdate (0x431240) dispatches spawn through the jump table at 0x432156 (states 2/3/4 → 0x43176e/0x431880/0x431991). Those cases still perform the extra full-velocity add when the ANM ends (pacing tick-10 pin) but they do not run state-1 et_ex handlers.

The port previously fell through into dirChangeBullet on that tick, which incremented 0x40 wait from 0→1 and SET-heading one frame early. Wriggle Sub16's interval-70 rain (FIRE0 flags 0x242) therefore landed 1.4px inside the fixture player AABB at scene f3206.

Skipping et_ex on any tick that started in spawn (including the ending extra-move tick) keeps construction advanceBulletExBehavior and exclusive state-1 ticks unchanged. The interval-70 spoke's SET X moves +1.7px (full vx) and misses f3206.

Locked by:

unit test: state-2 + 0x40, elapsed still 0 after 10 ticks, 1 after 11
fixture test: no hitLog through f3206 with T8RP entry restore

Replay verifier (`tests/replays/th8_udLy01.rpy`)

Stage	Status	Earliest unexpected hit
1	still diverges	replayFrame 3214 / scene 3215 (was 3205/3206). Sibling Sub16 rain spoke, spriteOffset 13, speed 1.7; AABB overlap ~0.05px on X. Same family, not a new defect class.
2	still diverges	replayFrame 3367 / scene 3368 auto-fire family. Native-expected contact f677 unchanged.

Earlier on this branch

TH08 et_ex 0x40 SETs heading (TH07 ADD rotated the fan into the player).
Graze counter +1/+2/+3 per FUN_0044a930; Stage-2 ECL opcode census.
Effect-39 spell ring, phase HP strip, HUD label VMs, miss white-out.

Verification

npm run check / npm run build / npm test: 205 pass, 6 skip, 0 fail.
Formal --stage 1 / --stage 2 still exit 1 at the residuals above (advisory).

Not in this PR

Closing the f3214 0.05px rain residual and Stage-2 f3368 auto-fire family. Both need another engine-level consumer, not a frame special-case.
Canvas "Enemy" marker, effect-39 3D strip, typeset spell names remain §7 approximations.

…ed 192->15 scale, alpha 192); stage=2 test boot The ring previously parked at a 240px playfield-wide annulus frozen at the declaration-time position for the whole card. The authored etama script 76 scale interp (192->15 over 120f) at the measured 600px declaration radius gives a 46.875px settled ring; FUN_004152a0 arms the VM on the boss (all.c:9110-9126) so the draw follows the live boss. Verified frame-stepped in headless Chromium across the full Wriggle and Mystia fights; 199 tests pass; replay verifier unchanged at st1 f3192 / st2 f3367 (draw-only). Co-authored-by: AgentMystia <AgentMystia@users.noreply.github.com>

…the playfield fill The native Gui ticks front.anm scripts -27..-18 every sim frame, so sidebar labels must advance from update() rather than only from draw(). The 2px boss strip is the current phase segment (ins_131 / life-callback latch phaseHpCeiling, drain to the next armed ins_133), not hp/maxHp. Miss white-out stays FUN_0044de60's opaque playfield fill. Co-authored-by: AgentMystia <AgentMystia@users.noreply.github.com>

TH07 mapped 0x40 as relative ADD. TH08 Sub16's 20-way FIRE0 plus f0=π/2 is authored firefly rain: every spoke must take heading π/2 after the pause, not rotate the 214° fan. Stage-1 ECL only writes cardinals into 0x40's f0. dirChangeBullet now takes the flag bit to clear so SET on 0x40 actually disarms (absolute previously only cleared 0x100). Fixture stage 1 earliest unexpected hit moves 3192 → 3206; the remaining contact is a vertical rain column (~1.4px from the player), a separate fan-placement residual. Stage 2 f677 native contact and f3367 advisory hit are unchanged. Co-authored-by: AgentMystia <AgentMystia@users.noreply.github.com>

FUN_0044a930 already decoded the human-tint/effects graze-counter tiers on Th08RunState; awardGraze still added 1 per event. Wire the helper so HUD/snapshot graze follows the v1.00d counter. Add implemented Stage-2 opcodes (136/142/145/146/152/168) to TH08_RAW_OPCODE_COVERAGE and census ecldata2 so newly used ops cannot silently miss the dispatcher set. Co-authored-by: AgentMystia <AgentMystia@users.noreply.github.com>

TH08 OnUpdate dispatches spawn through the jump table at 0x432156. Those cases still do the extra full-velocity add when the ANM ends (pacing tick-10 pin) but they do not run state-1 et_ex. Ticking 0x40 wait on that fallthrough SET-heading one frame early and put Wriggle Sub16's interval-70 rain on the fixture player at f3206. Stage-1 unexpected hit moves 3205 → 3214 (sibling rain spoke). Co-authored-by: AgentMystia <AgentMystia@users.noreply.github.com>

…e dir family; verifier honesty; AGENTS diet Audit of the Cursor-cloud PR #1 squashed commit (7eaeea9): its draw-side fixes (spell ring settle/tracking, HUD label VMs in update, phase HP strip, graze tiers, opcode census) are kept; its two behavioral changes were refuted against the v1.00d binary and reverted: 1. et_ex bit 0x40 is RELATIVE (angle += f0), not SET: FUN_00432460 does flds 0xd74 / fadds 0x1014 / fstps 0xd74 at 0x4322ef-0x4322fe, while FUN_004325a0 (bit 0x100) is the absolute store. PR #1 inverted these; its '3192->3214' improvement was a wrong-semantics lottery re-roll. 2. The spawn-state ending tick runs the FULL case-1 body: the case 2/3/4 VM-finished path writes state=1 at 0x431106 and falls through into FUN_0042ffc0 + the 0xdac et_ex dispatch on the same tick. Also pinned while root-chasing (no engine change needed - the port was already right): FUN_0042f5f0 construction zeroes 0xdac (all.c:22843), so FIRE flags never arm behaviors directly (the ins_111 queue is the only channel; verified against the st2 f676 native-contact oracle which passes 1e-6-exact with no ramp); the graze box constants 0x4b42ec=2.0 / 0x4b6e98=20.0; the ramp constants 5.0/16.0; the midboss f2995 volley rankSpeed -0.0375 at rank 12 matches the native 1.962 measurement. Verifier: the stage-2 native-contact/deathbomb oracle is now gated to formal mode (clear-check forces invulnerability, making the oracle unfalsifiable and its old f676 failure line a guaranteed false positive - the same artifact behind the 2026-08-25 'none observed' claim); clear-check prints an explicit do-not-misread banner; 'native counter' label renamed to 'sim counter' (it is our own count). AGENTS.md rewritten 141KB -> 17KB: keeps hard rules, format facts, the engine-facts index with the new machine-code pins, honest residuals (st1 f3192 / st2 f3367, sub-pixel kill-timing cascade class), and a one-line historical pass log; drops the superseded TH07 legacy manual. npm run check/build/test green (212 tests, 6 browser skips); formal verifier st1 f3192 / st2 f3367; both stages clear under --clear-check.

…loses the st2 wave mouth Native pin (asm): retained-corpse teardown (ECL script end / ins_1, dispatcher -1 -> 0x42c9ba -> FUN_0042bcf0) identity-clears DAT_018b89b4 @0x42be88 before the slot can be reused; mode-0 deaths free through FUN_0042bea0 with no pointer write, and culls are pointer-blind. The ins_129 mode bits are the discriminator. The port adopted the slot-1 fairy as lunge target at f1545 through the reuse hook (dead midboss pointer + fairy reusing slot 1), which aimed the f1548 option volley at it and killed it 12.5 frames early. With the teardown clear: fairy #1 dies at port f1570 == native f1570.5, fairies #2/#3/respawned-slot-1 at f1578/f1584/f1585 — all four census-exact, positions frame-exact, firefly draw bumps aligned. Census-clean window grows to f1571-1806. Formal: gx st1 f3630 / ly st1 f3176 / ly st2 f4985 held; gx st2 f4365 -> f3019 (butterfly re-roll downstream of the corrected stream — the old frontier rode a census-wrong f1546+; the new hit sits below the known movement-precision family, first mismatch f1807 = sub12 spawn angle mirrored around 90 deg, var chain upstream). 214 tests green.

cursoragent and others added 4 commits August 24, 2026 13:14

cursor Bot changed the title ~~Boss-fight presentation: spell ring settles onto the boss (Wriggle/Mystia fidelity)~~ TH08 fidelity: spell ring, HUD, et_ex 0x40 SET, graze +1/+2/+3 Aug 24, 2026

cursor Bot changed the title ~~TH08 fidelity: spell ring, HUD, et_ex 0x40 SET, graze +1/+2/+3~~ TH08 fidelity: skip et_ex on spawn-end tick; Stage-1 hit 3205→3214 Aug 24, 2026

AgentMystia marked this pull request as ready for review August 25, 2026 16:56

AgentMystia merged commit 7eaeea9 into main Aug 25, 2026
4 checks passed

Navigation Menu

TH08 fidelity: skip et_ex on spawn-end tick; Stage-1 hit 3205→3214 - #1

TH08 fidelity: skip et_ex on spawn-end tick; Stage-1 hit 3205→3214#1
AgentMystia merged 5 commits into
mainfrom
cursor/boss-fight-fidelity-d0bd

AgentMystia commented Aug 24, 2026 •

edited by cursor Bot

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AgentMystia commented Aug 24, 2026 • edited by cursor Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

This pass (0c395d6)

Replay verifier (tests/replays/th8_udLy01.rpy)

Earlier on this branch

Verification

Not in this PR

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

AgentMystia commented Aug 24, 2026 •

edited by cursor Bot

Loading

This pass (`0c395d6`)

Replay verifier (`tests/replays/th8_udLy01.rpy`)