Dead-air cleanup

Tighten without making it sound machine-edited, and only when the user asked for shortening.

Workflow

1. Silences: trimSilence on the clip uses its measured silenceRegions (measured for you when missing). Do not hand-pass guessed regions. On MCP, analyze_dead_air ranks removal ranges by confidence and already leaves about 120 ms of air each side; to reach the 150 to 300 ms below, shrink each range by a further 30 to 180 ms per side before cutting.
2. Filler and retakes: transcribe, then cutRanges with word-timed ranges in TIMELINE seconds, ripple:true so the gaps close.
3. Breathing room: leave 150 to 300 ms around speech; keep 250 to 400 ms between sentences and 600 to 900 ms before a punchline, reveal or emotional answer. Vary pause lengths: uniform crushing is the tell of an auto-editor.
4. Breaths are not filler. Keep them, especially before an emphatic line.
5. Hide the jumps by rotating techniques (punchIn, a cutaway, a 4 to 6 frame dissolve). The same punch-in twenty times is worse than the jump cuts.
6. Review every join with agentWorkspace action review mode speech (at most 30 s per request): no word crossing a cut, every thought complete, the ending makes sense.

Gotchas