fix: Untertitel ohne rollendes Fenster verlieren keine wiederholten Blöcke

vtt2txt.py entfernt Überlappungen nur noch, wenn die Datei ein rollendes
Fenster zeigt: mindestens die Hälfte der Blöcke wiederholt das Ende des
vorigen (YouTube-Auto-Untertitel: 97–99 %). Manuell erstellte Untertitel
werden vollständig übernommen, zwei gleiche Blöcke hintereinander bleiben
zwei.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

vtt2txt.py

@@ -19,6 +19,9 @@ SENTENCE_END = re.compile(r"[.!?…][\"'»«“”)\]]*(?=\s|$)")
 LINE_ENDS_SENTENCE = re.compile(r"[.!?…][\"'»«“”)\]]*$")
 
 MAX_SNAP_SECONDS = 30  # so weit darf ein Zeitraum bis zur nächsten Satzgrenze wachsen
+# Ab diesem Anteil von Blöcken, die das Ende des vorherigen wiederholen, gelten
+# die Untertitel als rollendes Fenster (YouTube-Auto-Untertitel: über 95 %)
+ROLLING_SHARE = 0.5
 TIME_EXAMPLES = "z. B. 00:05:10, 00:20 (Minute 20) oder 25 (Minute 25)"
 
 
@@ -204,21 +207,31 @@ def in_range(time, start, end):
     return (start is None or time >= start) and (end is None or time < end)
 
 
+def is_rolling(cues):
+    """True if most cues repeat the end of the previous one (rolling window)."""
+    pairs = list(zip(cues, cues[1:]))
+    overlapping = sum(1 for previous, cue in pairs if overlap(previous, cue))
+    return bool(pairs) and overlapping >= ROLLING_SHARE * len(pairs)
+
+
 def timed_lines(vtt):
     """Return (cue start, cue end, line) for the new caption lines of a VTT document.
 
     Auto captions show a rolling window: each cue repeats the end of the previous
     cue before adding new text. Only the lines after that overlap are new, so a
-    line the speaker really says twice in a row is kept.
+    line the speaker really says twice in a row is kept. Captions without a
+    rolling window (e.g. written by hand) are taken as they are, so identical
+    cues in a row stay separate.
     """
+    # Empty cues (YouTube transitions) say nothing about the window
+    cues = [(start, end, cue) for start, end, cue in read_cues(vtt) if cue]
+    rolling = is_rolling([cue for _, _, cue in cues])
     lines = []
     previous = []
 
-    for start, end, cue in read_cues(vtt):
-        # Empty cues (YouTube transitions) say nothing about the window
-        if not cue:
-            continue
-        lines.extend((start, end, line) for line in cue[overlap(previous, cue):])
+    for start, end, cue in cues:
+        new = cue[overlap(previous, cue):] if rolling else cue
+        lines.extend((start, end, line) for line in new)
         previous = cue
 
     return lines