dasVideo

Living checklist. Frozen design rationale in ORIGINAL_PLAN.md. No time/effort sizing — it takes what it takes.

Legend: [ ] todo · [~] in progress · [x] done

P0 — Scaffold + CI (foundation; green from commit one)

• CMake linking daslang via DASLANG_DIR (dasVulkan-style); .shared_module output
• .das_module (loads dasModuleVideo.shared_module; needs require daslib/fio)
• vendor/pl_mpeg.h (MIT, compiled into the module TU)
• LICENSE (MIT), README.md, .gitignore, .gitattributes
• CLAUDE.md (repo instructions for AI collaboration)
• doc/ Sphinx skeleton
• empty module compiles + headless smoke passes (proven locally on Windows)
• [~] CI: Linux + macOS build + headless smoke (regular GitHub runners). Windows CI is a fast-follow — the module build is proven locally; the daslang Windows build still needs bison/flex + MSVC env wiring on windows-latest.

P1 — pl_mpeg video (headless)

• video_open → VideoPlayer handle (opaque); info via getters (width/height/framerate/samplerate/duration/frame_time, y/uv strides). A plain frame-info struct is ergonomic sugar deferred to the daslib boost layer.
• video_decode → native YUV420p planes (borrow-until-next-decode)
• get_data block accessor — YUV planes (zero-copy array<uint8># views)
• C++ YUV→RGB converter (plm_frame_to_rgba); get_data — RGBA8
• [~] headless decode-N-frames + checksum test (green: local Windows + CI Linux/macOS; Windows CI shares P0's fast-follow)

P2 — GL output

• examples/play_rgba_gl — RGBA → one texture, fullscreen quad; clargs CLI, loop-by-default (--noloop/--max-frames/--scale/--shot). Visually verified (test pattern renders in a window). Surfaced + fixed an opaque-alpha bug in the RGBA converter along the way.
• examples/play_yuv_gl — 3×GL_R8 textures + YUV→RGB fragment shader (BT.601, matches pl_mpeg). Per-plane GL_UNPACK_ROW_LENGTH for the macroblock-rounded strides; @stage = 0/1/2 to bind the 3 samplers to distinct units. Visually verified against the RGBA path.

Follow-ups

• daslang teardown crash (was: clargs early-exit like --help + the GL stack → g_envTotal=1). Root cause: defer-lowered finally ran on the early return and read a finally-referenced local whose stack slot the interpreter's reuse optimization had handed to an earlier sub-scope, clobbering its zero → garbage pointer → AV. Fixed upstream in daslang (disable slot reuse for blocks with a finalList) + regression test. No dasVideo change was needed.

P3 — Audio + sync

• decode MP2 → interleaved float PCM (video_decode_audio + get_audio_data borrow; audio is opt-in via video_enable_audio so video-only paths stay zero-overhead)
• play_sound_from_pcm_stream + append_to_pcm top-up loop (feed_audio, AUDIO_TARGET_CHUNKS slack)
• consumed_position master clock; stream_que_length backpressure
• pace loop: top-up / promote-when-due / drop-intermediate; clean looping via drain-then-restart (re-anchor the clock so A/V stays in sync across the boundary)
• silent / no-audio wall-clock fallback (--mute + auto when no audio stream)
• examples/play_audio_gl — windowed A/V-synced player (real playback verified: consumed_position advances, video frames pace to it)
• test with an audio-bearing clip (tests/test_audio.das + test_audio.mpg fixture; device-free decode+borrow, wired into CI)

Demo presenter — Emma

• Emma — synthetic face + voice (Kokoro bf_emma → Sonic talking-head). The full reproducible recipe is in EMMA.md; canonical source faces in assets/emma/. assets/emma/emma_intro.mpg (1024 master → 384×384 MPEG-1/MP2) is play_audio_gl's default, so the audio demo is "Emma introduces dasVideo" with real A/V sync.

Plan order (per Boris): docs/tutorials → daspkg → AV1 after. (P4 imgui later.)

P6 — Docs + tutorials ← NEXT

• Sphinx reference content (per-API) for the current surface (open/decode/get_data/ get_audio_data/the audio-sync model)
• tutorial ladder: open-probe → rgba-to-gl → yuv-shader → audio-sync → (imgui-embed) → (av1)
• recordings (voiced, self-verifying) per the standard pipeline — Emma fronts them

daspkg — register (unblocked: it's demoable now)

• daspkg-index entry; build() → cmake_build() sets DASLANG_DIR

P4 — dasImgui (later)

• image() embed in an imgui window — flagship demo

P5 — dav1d backend (the AV1 follow-up, after docs/daspkg)

• vendor dav1d; WebM demux; Opus/Vorbis audio decode
• behind the same VideoSource API; backend selection by probe/extension

Cross-cutting

• CI stays green every phase; //! docstrings written with the code
• daspkg-index entry (publish once demoable)

Deferred

• WASM target — BLOCKED: daslang WASM backend lacks OpenGL + dasAudio. Revisit after that gap is closed upstream.