fix false gap regions: close cracks between adjacent seg_windows

The gap detection was firing on cracks between consecutive segment windows
where unmatched (interpolated) segments sat between two matched anchors.
Whisper words belonging to real content fell in the crack and were
mis-classified as out_of_transcript.

Root cause: seg_window[i].end < seg_window[i+1].start whenever the matched
segment's LCS span ended before the next segment's start position. The 8s
MIN_AD_GAP threshold was then applied to this crack, producing fake 'gaps'.

Fix: after building all seg_windows, extend each window's end to the next
window's start (max of current end and next start). The gap detector then
only sees spans genuinely not covered by any window.

Impact on ep152:
  Before: 88.6% assigned, 6 gap regions (5 fake + 1 real)
  After:  99.8% assigned, 1 gap region (real StepStone ad, 36s intro)

Author: Zenflow

scripts/podcast_to_moshi_dataset.py

@@ -546,6 +546,20 @@ def min_score(n: int) -> float:
         re = seg_raw_end[si]   or 0.0
         seg_windows.append((rs, re, seg.get("speaker", ""), seg.get("text", "").strip()))
 
+    # ── Close cracks between adjacent windows ────────────────────────────────
+    # After interpolation, consecutive segment windows may not be contiguous:
+    # the end of window[i] < start of window[i+1].  Any whisper word that
+    # falls in this crack gets misclassified as out_of_transcript even though
+    # it belongs to real content.  Extend each window's end to the next
+    # window's start so cracks are closed.  The gap-detection below then only
+    # sees spans that are genuinely not covered by any window.
+    closed = []
+    for i, (rs, re, spk, txt) in enumerate(seg_windows):
+        if i + 1 < len(seg_windows):
+            re = max(re, seg_windows[i + 1][0])
+        closed.append((rs, re, spk, txt))
+    seg_windows = closed
+
     # ── Detect real gaps ≥ MIN_AD_GAP ─────────────────────────────────────────
     gap_regions = []
     if seg_windows[0][0] > MIN_AD_GAP: