Tighten a recording

1. The recording must be ingested and transcribed (ingest, then transcribe; ask before any whisper model download). Use its asset id.
2. Dry run: tighten {project_dir, asset}. Show the user the summary: total length before → after, pauses shortened, and every filler and retake cut with its time and words. Also show the join check (joins): every partial_word (a cut inside a word) and repeated_word (the same word on both sides of a join) with its time and fix (the cut moved to the nearest word gap).
3. Let the user adjust before applying:
   • "keep the pauses" → silences: false, or a gentler max_pause_ms: 1200
   • "pace it like my other videos" → pacing_from: an analyze result in the project (qa/analysis.json, from a video the user edited) or the asset id of such a video. It sets max_pause_ms to their p95 pause (clamped to 250–1500 ms) and keep_pause_ms to their median pause (120–600 ms); explicit values win. Report the numbers (pacing).
   • "don't cut retakes" → retakes: false, if a "false start" was intentional
   • a retake detection that is wrong: turn retakes off; never hand-edit cuts
4. Apply: tighten {project_dir, asset, …same options, apply: true}. It writes a new asset <asset>-tight (the original is kept) with the transcript re-timed, so captions and claim_refs line up with the new cut.
   • Apply refuses while a partial_word remains. Usually the transcript's timings are off there: re-transcribe, or change the options. Pass force: true only when the user has heard the join and accepts it.
   • With whisper installed, apply re-transcribes ±2 s around every join (asr_check): report each join_mismatch (expected vs heard words) so the user can listen to it. not_run gives the reason (no whisper, no model); it is not an error.
5. Next: shorts or a talking-head / silent-vlog spec on <asset>-tight. Suggest watching the tightened clip once: word-timing cuts are precise to whisper's timestamps, and a rare cut can clip the edge of a word that the join check cannot see. For a video, the compare skill builds a before/after page for that: a: {file: <the original's project-relative path>} against b: {file: <the path tighten returned, source/assets/<asset>-tight.mp4>}, played in sync, so the user can hear each cut against the original.