feat(backend): demucs two-stem -> mp3 adapter

separate.py

@@ -0,0 +1,42 @@
+import subprocess
+import sys
+from pathlib import Path
+from processor import _get_device
+
+
+def _run_demucs(audio_path: Path) -> Path:
+    """Run demucs in two-stems mode targeting vocals; return path to no_vocals stem."""
+    out_root = audio_path.parent / "demucs_out"
+    out_root.mkdir(parents=True, exist_ok=True)
+    cmd = [
+        sys.executable,
+        "-m",
+        "demucs",
+        "--two-stems", "vocals",
+        "-n", "htdemucs",
+        "--device", _get_device(),
+        "--mp3",
+        "--mp3-bitrate", "192",
+        "-o", str(out_root),
+        str(audio_path),
+    ]
+    subprocess.run(cmd, check=True)
+    # demucs writes to {out_root}/htdemucs/{audio_stem}/no_vocals.mp3
+    candidate = out_root / "htdemucs" / audio_path.stem / "no_vocals.mp3"
+    if not candidate.exists():
+        raise RuntimeError(f"demucs did not produce {candidate}")
+    return candidate
+
+
+def _encode_mp3(wav_path: Path, mp3_path: Path) -> Path:
+    """Re-encode (or copy) a stem file to the target mp3_path."""
+    subprocess.run(
+        ["ffmpeg", "-y", "-i", str(wav_path), "-b:a", "192k", str(mp3_path)],
+        check=True, capture_output=True,
+    )
+    return mp3_path
+
+
+def separate_vocals_to_mp3(audio_path: Path) -> Path:
+    no_vocals_stem = _run_demucs(audio_path)
+    return _encode_mp3(no_vocals_stem, audio_path.parent / "no_vocals.mp3")

tests/test_separate.py

@@ -0,0 +1,21 @@
+from unittest.mock import patch, MagicMock
+from pathlib import Path
+from separate import separate_vocals_to_mp3
+
+
+@patch("separate._run_demucs")
+@patch("separate._encode_mp3")
+def test_separate_returns_mp3_path(mock_encode, mock_demucs, tmp_path):
+    audio = tmp_path / "in.m4a"
+    audio.write_bytes(b"x")
+    no_vocals_wav = tmp_path / "no_vocals.wav"
+    no_vocals_wav.write_bytes(b"y")
+    mock_demucs.return_value = no_vocals_wav
+    out_mp3 = tmp_path / "no_vocals.mp3"
+    mock_encode.return_value = out_mp3
+
+    result = separate_vocals_to_mp3(audio)
+
+    assert result == out_mp3
+    mock_demucs.assert_called_once_with(audio)
+    mock_encode.assert_called_once_with(no_vocals_wav, audio.parent / "no_vocals.mp3")