10-bit HDR mixer: SDR/HLG/PQ conversion, native 4:2:2, HDR10 outputs, shutdown fixes

[jpietek]

10-bit HDR mixer: per-source SDR/HLG/PQ conversion, native 4:2:2 canvases, HDR10 outputs, and the shutdown fixes.

Stacked on #34 (FFmpeg 8.0/8.1 series); merge #34 first, then this PR (squash).

What it does

• Colour: every source is normalised to the canvas contract (sdr, hlg, pq) on the GPU by the new tonemap_cuda filter (auto per-frame resolution, untagged files are BT.709 SDR); renditions convert back with clip/mobius/hable, a knee (tonemap_param) keeps SDR content exact inside an HDR canvas.
• 10-bit / 4:2:2: P010 and P210 canvases, 8-bit sources promoted in the compositor, v210 sources (packed 10-bit 4:2:2) stay 4:2:2 through conversion and compositing; tonemap_cuda reads and writes NV12/P010/NV16/P210 and resamples in one pass.
• Outputs: HEVC Main10 with HLG/PQ VUI over Janus and files; PQ carries HDR10 static metadata (enc_video hdr_metadata, values from pyplumber.mixer.color). Needs nv-codec-headers 13 (driver ≥ 570), both Dockerfiles updated.
• Runtime fixes: filter node no longer drops queued frames when an EOF marker arrives mid-wait (multi-input transitions lost their tail); Python shutdown deadlock and stopped-queue peeks fixed (from codex/recorder-shutdown, folded in), with the seek-flush latch cleared so seeks/speed changes keep flowing; per-node control bindings release the GIL.
• Code layout: pure code motion plus libav/avcpp reuse. Compositor: src/nodes/hwaccel/cuda_rect_draw.{hpp,cpp} next to a ~700-line cuda_rect_overlay.cpp. Mixer control: src/mixer/orchestrator/ (MixerOrchestrator.hpp, core.cpp, one file per transition kind), src/mixer/primitives/ (CamelCase component headers: TickGrid over av::Rational, Cadence, MonotonicClock, CutLatency, Snapshot, MixerState, ...; snake_case free-function modules: pixel_layout, compositor_layers, compositor_geometry), plus routing.hpp, graph_ops, TransitionScheduler, Playout.hpp. Hand-rolled rescales, plane maths, HDR side-data sizes and colour ints replaced by av_rescale*, av_inv_q, av_image_get_linesize, av_pix_fmt_get_chroma_sub_sample, the libavutil metadata allocators and AVColorTransferCharacteristic; AGENTS.md now asks for that. Fix: source_switcher no longer queries the timeline at AV_NOPTS_VALUE. Layout table in doc/mixer.md; new CPU unit tests for pixel layout, HDR10 metadata serialisation, canvas.latency_ms, and an SDR media wipe on an HLG P210 canvas; one cpp_binary fixture builds the standalone C++ tests.
• Python package: avpmixer/ becomes pyplumber/mixer/, so images ship one Python package; the package root imports the native module lazily so configuration, control and color helpers load without it.
• Demo tooling: make_config.py emits HDR shows (--color, --working-format, --sdr-port, colour-tagged and v210 sources); tests/sdr_patterns.py renders SDR pattern clips. README shows the 16-source 1080p60 HDR example.
• Config/CLI: working_format, color, latency_ms, v210 sources, rendition color/tonemap*/max_cll/max_fall; --input-color, --working-format; only semiplanar working formats; 32-source cap enforced at load; docs/config.md opens with a complete loader-validated example, plus a shipped HDR example.

Known limitations (documented in demos/mixer/docs/config.md)

32 sources per show (compositor pad mask), all sources always running, no runtime source add/remove, --input >32 files needs homogeneous inputs, NVENC output is 4:2:0, HLG/PQ need a 10-bit canvas.

Verification (all on this head or its immediate predecessors, see commits)

• Unit: mixer 114, colour 22, replay 76, playlist 73, playout 25, dmabuf 5; after the module split and libav reuse: 135 CPU tests, the four GPU smokes, check_transition_recovery (plain and --prewarm), and the live 16-source HDR show (cuts, fades, media wipes on the HLG canvas, both Janus legs fresh) all green on the final head.
• GPU smokes (T4): v210, interop matrix, 10-bit mixer incl. 16-source grids, transfers matrix (3×3 transfers × depths, round trips, rejections), output qualification (HEVC Main10 HLG VUI, PQ + HDR10 SEIs, SDR round trip within 7 codes), shutdown smoke, live takes (24 cuts/fades on a mixed 8/10-bit show, both Janus legs fresh).
• Demos: cuda-overlay matrix 12/12, replay integration 24 fps module 37 passed, playlist scripted run, dmabuf-browser unit tests.
• Downstream: eka-recorder e2e PASS ×2 (tests/downstream/eka-recorder-smoke.sh, 0 panics; the Sep-15 baseline on the same variant panicked); sports-live-reframer legacy smoke passes (v2 variant has a pre-existing, unrelated node-factory failure, identical on Sep 15).
• Regression found and fixed during review: replay speed-continuity failure caused by the latched queue-stop flags after a seek (A/B confirmed), fixed in graph_core.hpp.

Follow-ups (not in this PR)

Router-first topology (per-slot conversion, runtime source add/remove, lifts the 32-source cap); reframer/recorder repos need the FFmpeg 8 migration (patch dir, avcpp pin, band-blur patch port); Dockerfile.fedora is broken on develop already (no pkg-config for FFmpeg headers).

🤖 Generated with Claude Code