fix(discovery): fall back to directory name for audio model type detection

[ryancee]

Summary

Models whose config.json omits a top-level model_type field were being classified as llm by detect_model_type(), causing a KeyError('model_type') crash at inference time.

Root Cause

detect_model_type() in model_discovery.py identifies audio models exclusively by matching config["model_type"] and config["architectures"] against the mlx-audio model-type sets. Models using NeMo-format config files — notably parakeet-tdt models — do not include a top-level model_type field. This causes:

1. detect_model_type() falls through all audio checks and returns "llm"
2. A BatchedEngine (LLM engine) is loaded for the model
3. The LLM loading path calls config["model_type"] without a .get() fallback
4. KeyError: 'model_type' is raised — reported to the client as HTTP 500

POST /v1/audio/transcriptions → 500: 'model_type'

The model itself works fine when loaded directly via mlx_audio.stt.utils.load_model, which already has its own directory-name fallback in base_load_model.

Fix

After all config-based checks, extract the first hyphen-separated segment of the model directory name and match it against AUDIO_STT/TTS/STS_MODEL_TYPES — the same sets already used for config-based detection. The segment is guarded against _LLM_TYPE_COLLISIONS to prevent false positives on directories named llama-..., qwen3-..., etc.

# e.g. "parakeet-tdt-0.6b-v2" → stem = "parakeet" → matches AUDIO_STT_MODEL_TYPES → "audio_stt"
dir_stem = model_path.name.lower().split("-")[0]
if dir_stem and dir_stem not in _LLM_TYPE_COLLISIONS:
    if dir_stem in AUDIO_STT_MODEL_TYPES:
        return "audio_stt"
    ...

This mirrors the existing logic in mlx_audio.utils.base_load_model, which already resolves model type from the directory name when config.json is missing the field.

Affected Models

Any audio model with a NeMo-format config, including:

• parakeet-tdt-0.6b-v2 (confirmed: was returning HTTP 500 before fix, now correctly detected as audio_stt)
• Any future parakeet or NeMo-converted model added to ~/.omlx/models/

Note

A companion PR to Blaizzy/mlx-audio#657 adds model_type: "parakeet" injection in ModelConfig.from_dict(), which addresses the root cause at the mlx-audio level. This omlx fix provides defense-in-depth for any other audio models that may have the same omission.