Stream capture frames directly to FFmpeg

[abrichr]

What changed

• stream RGB capture frames directly from memory into a separately provisioned FFmpeg process
• deterministically fill missing integer PTS slots without relying on wall-clock scheduling
• embed the logical capture-frame map in a valid MP4 UUID box so nearest-frame extraction keeps its existing semantics
• bound encoder input with a queue and fail loudly on stalls, short writes, broken pipes, close failures, timeouts, or encoder errors
• retain verification and atomic promotion while eliminating PNG and ffconcat staging

Why

Capture 1.0.3 removed PyAV from the runtime boundary but temporarily staged every frame as a PNG before final MP4 encoding. That added disk I/O, delayed finalization, and duplicated work. This restores the efficient direct-to-video architecture without bundling codec libraries into the MIT package.

User impact

Recording writes compact video while it runs. Stop/finalization no longer performs a second full image-sequence encode, and no transient screenshot sequence is left on disk. Timestamp lookup remains deterministic and compatible with the compiler.

Validation

• independent audit: no merge-blocking findings
• full suite: 324 passed, 7 skipped
• focused video suite: 28 passed, including real FFmpeg/ffprobe timing, backlog independence, nearest-frame lookup, and faststart remux
• adjacent capture/runtime suite: 115 passed, 1 skipped
• Ruff and diff checks pass
• wheel and sdist archive-boundary validator passes