cleanup: fix fidelity comment, rename opaque var, drop MMT-repo reference

.gitignore

@@ -2,5 +2,4 @@ __pycache__/
 *.pyc
 .DS_Store
 .venv/
-MMT-repo/
 _outputs/

app.py

@@ -49,7 +49,7 @@ def load_model():
     """Build the model on CPU and load the sod/ape checkpoint weights. Stays on
     CPU here even if GPU hardware is available -- ZeroGPU only allows touching
     CUDA from inside an @spaces.GPU call, not at module load time."""
-    m = music_x_transformers.MusicXTransformer(
+    model = music_x_transformers.MusicXTransformer(
         dim=train_args["dim"],
         encoding=encoding,
         depth=train_args["layers"],
@@ -65,9 +65,9 @@ def load_model():
     state_dict = torch.load(
         CKPT_DIR / "checkpoints" / "best_model.pt", map_location="cpu"
     )
-    m.load_state_dict(state_dict)
-    m.eval()
-    return m
+    model.load_state_dict(state_dict)
+    model.eval()
+    return model
 
 
 # Checkpoint is small (~80MB) and lives in the Space repo, so load it
@@ -224,7 +224,9 @@ def process_fn(input_midi_path, instruments_text, generation_length, temperature
     audio_path = None
     if render_audio:
         audio_path = get_default_path(ext=".wav")
-        music.write(audio_path)  # uses the MuseScore General soundfont muspy fetches on first use
+        # Uses the MuseScore General soundfont muspy fetches on first use. The original repo's
+        # polyphony option isn't supported by muspy==0.5.0's write_audio() (no passthrough kwarg).
+        music.write(audio_path)
 
     report_path = get_default_path(ext=".txt")
     Path(report_path).write_text(f"{report_name}\n\nInstrument Matching\n{instrument_note}\n")