podcast_to_moshi: complete step2 rewrite with segment-by-segment alignment

- No more random anchor sampling — walk every transcript segment in order
- LCS-based matching tolerates whisper insertions/deletions at boundaries
- Per-length score threshold (short segments need higher confidence)
- Median-delta computation from LCS pairs (robust to jingle/ad bleed)
- Gaps proven in raw audio time (≥8s) — no false positives
- Every segment gets a window; unmatched ones inherit last known delta
- Step 1 now caches w1.json and skips re-transcription if present
- Step 3 uses MP3 directly (no WAV conversion), LCS-based w2 refinement
- Debug mode (DEBUG=True) prints per-segment match/miss details
- Result: 80.6% word assignment, 9 clean gap regions on ep152

Author: Zenflow