[codeonym]

Summary

Record now turns narration into local Qwen3-TTS 0.6B speech using Ryan in English, with pinned offline assets and the official CPU runtime.

capture → close browser/display → synthesize narration
  → demo.audio/NNN.wav
  → transcript voice, duration and SHA-256 metadata

Blank narration has no clip. Reruns replace this demo's generated clips and preserve unrelated files. CI has a 45-minute budget for image building and CPU synthesis.

Evidence

• Before: the new recording test failed because voice metadata was absent.
• After: type checking, build, 50 unit tests and all 10 Docker E2E checks pass in one clean run (968 seconds).
• Fresh network-disabled processes produce identical WAV hashes, including repeated text at different indexes. Final recording step starts differ by 2, 2 and 3 ms; existing caption playback and frame-alignment checks pass.
• Missing assets fail the worker and recording command without publishing a success transcript. Sample narration was listened to and approved.

Merge Danger

Door: two-way

Blast Radius: recording

The image is about 7 GB; observed synthesis peaks at about 5.6 GiB of process memory and takes minutes on the tested CPU. MP4 remains silent until GH-5 adds holds and muxing. Schema version 1 and recorded timing are preserved. The existing odd-viewport encoder hang is recorded as a separate approved follow-up.

Closes #4