`logs/session-debrief/_work/narr-t4.txt`

1 lines
Fourth task. The hardest one. The narration was a single audio track but the video was choreographed around per-task windows. They drifted apart. So we refactored the entire audio pipeline. Each task gets its own narration MP3, measured exactly. We compute a schedule. We pad every clip with silence to fill its visual window. The schedule gets embedded in the ledger so the visual choreography knows the exact frame each card should appear. And we caught a nasty ffmpeg bug where the apad filter without a length cap produces infinite silence — the build would have run for hours.