align LRC from ready text.txt via stable-ts and auto-detect language

- add align_text_to_audio(): forced alignment without ASR, fast
- add _StableTSModels cache; detect language via stable-ts when language=None
- LRC stage: use existing text.txt, copy it if already LRC, fall back to ASR only
- filter junk lyrics from tags and ASR hallucinations (promo phrases, tics)
- mark empty ASR results with metadata.empty_text and skip them on resume
- warn in GUI when LRC is run with transcription disabled
- add mark_no_speech.py one-shot script for existing tracks

mark_no_speech.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+"""Mark tracks with metadata.empty_text=true if their text is missing or hallucinated."""
+import json, sys
+from pathlib import Path
+
+HALLUCINATION = (
+    "продолжение следует", "субтитры сделан", "субтитры создава",
+    "created with", "russian-area.de", "музыка музыка",
+)
+
+
+def main(root: Path) -> int:
+    count = 0
+    for meta in root.rglob("metadata.json"):
+        track = meta.parent
+        txt = track / "text.txt"
+        if txt.is_file():
+            body = txt.read_text(encoding="utf-8", errors="ignore").strip().lower()
+            # If the text is real (not hallucination) and not empty, skip marking.
+            if body and not any(h in body for h in HALLUCINATION):
+                continue
+            # Remove the text.txt so the next run will re-process it.
+            txt.unlink()
+        # Set empty_text=True in metadata.
+        try:
+            data = json.loads(meta.read_text(encoding="utf-8"))
+        except (OSError, ValueError):
+            data = {}
+        data["empty_text"] = True
+        meta.write_text(
+            json.dumps(data, indent=2, ensure_ascii=False), encoding="utf-8",
+        )
+        print(f"marked: {track}")
+        count += 1
+    print(f"\nmarked {count} tracks")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(Path(sys.argv[1] if len(sys.argv) > 1 else ".")))