[codeonym]

Summary

Short actions previously ended before their separate speech clips, and the delivery MP4 was silent. Capture actions once, synthesize afterward, then fit and mux narration on a shared timeline.

capture once → synthesize → plan frame holds → retime master/events/captions
                                         → normalize voice → mux AAC → validate sync
host compare → validate transcripts → compare content and starts within 150 ms

Preserve original WAVs/hashes and calibration pre-roll. Blank-only demos stay video-only.

Closes #5

Evidence

• Before: silent MP4s and step windows shorter than speech.
  After: narrated MP4s with at least 200 ms speech tails; playback approved.
• npm test: 63 passed. Typecheck, build, package dry run and whitespace checks pass.
• All 19 media checks verified. The full run passed the non-recording files; after a capture overload failure, the unchanged recording file passed all 9 tests on retry with free CPU.
• Final step-start drift: 5 / 40.333 / 9 ms. Voice: -16.42 LUFS / -1.98 dBTP. Click discrepancy: -12.493 ms, within two frames.
• Standards and Spec reviews: zero unresolved findings.

Merge Danger

Door: two-way

Blast Radius: recording

Changes final media timing and adds host transcript comparison. The existing capture-overload guard remains unchanged; recording still needs spare CPU. Comparison reads transcript metadata only.