test(music): add objective vocal-alignment probe to the A/B harness

Force Romanian Whisper transcription of each rendered test song and count
expected Moldovan words. Turns the A/B from a listening impression into a
measurable signal: MiniMax 0/4 (vocalised as hums at 15s), YuE2-3B 2/4,
ACE-Step 1/4.

scripts/compare_music_tests.py

@@ -79,6 +79,25 @@ def main() -> int:
         print(f"\n[{len(rows)}/{len(MODELS)} models present] Verdict blocked until the quota ladder renders the rest.")
         return 2
 
+    # Objective lyric-alignment probe: force Romanian and count expected words.
+    # Uses faster-whisper when importable (scratch venv), else skips cleanly.
+    try:
+        from faster_whisper import WhisperModel  # type: ignore
+
+        wm = WhisperModel("small", device="cpu", compute_type="int8")
+        expected = ["liniste", "ploua", "paine", "poiana"]
+        print("\n=== Vocal lyric alignment (Whisper small, language=ro) ===")
+        for slug, label in MODELS.items():
+            p = base / f"{slug}_test.wav"
+            if not p.exists():
+                continue
+            segs, _ = wm.transcribe(str(p), language="ro", beam_size=5)
+            txt = " ".join(s.text.strip() for s in segs).lower()
+            hits = sum(1 for w in expected if w in txt)
+            print(f"{label:28} ro-match {hits}/4 | {txt[:150]}")
+    except ImportError:
+        print("\n(faster-whisper not importable; skipping vocal alignment probe)")
+
     # Lyric-quality judgement is human/LLM-eval; print the checklist template.
     print("\n=== Lyric-quality A/B checklist (Moldovan lines: 'Liniste, liniste, afara ploua...') ===")
     for line in [