fix: is_valid_module_name rejects names containing '.'

[kmosher]

Summary

is_valid_module_name currently only validates the first character of a candidate, so it accepts strings that contain . (e.g. parakeet_tdt_0.6b). get_model_category passes such candidates straight into importlib.util.find_spec(f"mlx_audio.{category}.models.{candidate}"), which interprets the dot as a package separator and tries to import a parent module that doesn't exist — raising ModuleNotFoundError and aborting model resolution.

The result: every model whose repo name contains a . (which is nearly every parakeet variant — *-0.6b-*, *-1.1b-* — and most Whisper variants like whisper-large-v3) crashes during load, even though the parakeet / whisper candidate would have resolved correctly on a later iteration.

Reproduction

from mlx_audio.utils import load_model
load_model("mlx-community/parakeet-tdt-0.6b-v3")

Traceback (most recent call last):
  File ".../mlx_audio/utils.py", line 778, in get_model_category
    if importlib.util.find_spec(module_path) is not None:
  File ".../importlib/util.py", line 94, in find_spec
    parent = __import__(parent_name, fromlist=['__path__'])
ModuleNotFoundError: No module named 'mlx_audio.tts.models.parakeet_tdt_0'

The candidate parakeet_tdt_0.6b (synthesized by get_model_name_parts joining parakeet, tdt, and 0.6b with _) is passed to find_spec which splits on . and looks for mlx_audio.tts.models.parakeet_tdt_0.

Fix

Replace the partial first-character check in is_valid_module_name with Python's canonical identifier check, str.isidentifier(). This correctly rejects any string containing . (or other non-identifier characters), so get_model_category's loops skip the dotted candidate and continue to the valid parakeet / whisper hint.

 def is_valid_module_name(name: str) -> bool:
     """Check if a string is a valid Python module name."""
     if not name or not isinstance(name, str):
         return False
 
-    return name[0].isalpha() or name[0] == "_"
+    # Must be a valid Python identifier: letter/underscore start, then
+    # only letters/digits/underscores. Crucially rejects names containing
+    # '.' (e.g. "parakeet_tdt_0.6b"), which importlib.util.find_spec would
+    # otherwise interpret as a package path and crash with
+    # ModuleNotFoundError on the synthetic parent.
+    return name.isidentifier()

After the fix, the same load_model("mlx-community/parakeet-tdt-0.6b-v3") call resolves via the parakeet candidate → mlx_audio.stt.models.parakeet, downloads weights, and serves transcription requests normally.

Test plan

• Added mlx_audio/tests/test_utils.py::test_is_valid_module_name_rejects_dots covering the regression and a handful of valid/invalid identifier cases.
• Manually verified end-to-end: mlx_audio.server running on :8890, POST /v1/audio/transcriptions with model=mlx-community/parakeet-tdt-0.6b-v3 and model=mlx-community/parakeet-tdt_ctc-110m both load and transcribe successfully.
• Existing tests pass (uv run pytest mlx_audio/tests/).

Affected models (non-exhaustive)

Any HuggingFace repo whose name contains ., including:

• mlx-community/parakeet-tdt-0.6b-v3, -v2
• mlx-community/parakeet-ctc-0.6b, parakeet-rnnt-0.6b, parakeet-tdt-1.1b, parakeet-rnnt-1.1b, parakeet-ctc-1.1b
• mlx-community/parakeet-tdt_ctc-0.6b-ja, parakeet-tdt_ctc-1.1b