Speaker Diarization Cleanup

Auto-diarization has two predictable failure modes that a post-processing pass can fix without retraining: short spurious swaps (easily filtered by duration) and generic speaker labels (easily fixed by matching against known voices). Both are cheap to detect and apply.

1. Parse the diarized transcript into turns. 2. Merge turns shorter than a configurable threshold (default 2s) into the neighboring dominant speaker. 3. For each anonymous speaker label, compute an average voice embedding from their turns. 4. Compute embeddings for each reference clip. 5. Match anonymous labels to reference labels by cosine similarity above a threshold; unmatched speakers keep their numeric label. 6. Emit the cleaned transcript + a diff showing what changed.