Deepgram word-level timestamps = frame-accurate sentence cuts

When transcribing podcast or video audio for clipping, Deepgram Nova-3 returns WORD-LEVEL timestamps, not just sentence-level. Pull them with `&utterances=true&punctuate=true`. Each word gets start and end in ms. Makes frame-accurate sentence cuts trivial: cut at the last word's end, resume at the next sentence's first-word start. No guesswork, no breath clips hanging on front or back.