Moss tts server

[voxmenthe]

Server API Surface: Request Fields, Reserved model_kwargs Safety, Optional Fallbacks

This PR focuses on server-facing request handling and validation behavior for audio endpoints, with emphasis on safe passthrough semantics and clear failure modes.

Scope (Included)

This PR includes:

• mlx_audio/server.py
• mlx_audio/tests/test_server.py
• mlx_audio/sts/voice_pipeline.py
• mlx_audio/sts/tests/test_voice_pipeline.py
• mlx_audio/stt/models/qwen3_forced_aligner/__init__.py

Primary behavior in scope:

• OpenAI-compatible /v1/audio/speech request-surface expansion for advanced generation controls.
• Validation guard for reserved model_kwargs collisions to prevent ambiguous/duplicate generation arguments.
• Explicit HTTP 400 response for reserved collisions with clear error detail.
• Server-side passthrough hygiene for strict model signatures.
• Optional dependency fallback behavior on server/STS paths covered by tests.

Scope (Intentionally Excluded)

This PR does not include:

• Realtime runtime internals in mlx_audio/tts/models/moss_tts_realtime/**.
• Non-realtime TTS runtime internals in mlx_audio/tts/models/moss_tts/**.
• General CLI gateway behavior in mlx_audio/tts/generate.py.
• Broad docs/readme/example expansion.

Validation

uv run pytest -q \
  mlx_audio/tests/test_server.py \
  mlx_audio/sts/tests/test_voice_pipeline.py

Targeted safety check executed:

uv run pytest -q mlx_audio/tests/test_server.py -k tts_speech_rejects_reserved_model_kwargs

Result: 1 passed.

Manual Verification Notes

• Reserved model_kwargs collisions return HTTP 400 and include clear guidance:
  • Server validation path: _validate_reserved_model_kwargs(...) in mlx_audio/server.py.
  • Error detail includes:
    • model_kwargs contains reserved generation arguments (...)
    • Pass these as top-level request fields instead.
• Confirmed PR diff scope is server-focused with no model-runtime internal files:
  • mlx_audio/server.py
  • mlx_audio/tests/test_server.py
  • mlx_audio/sts/voice_pipeline.py
  • mlx_audio/sts/tests/test_voice_pipeline.py
  • mlx_audio/stt/models/qwen3_forced_aligner/__init__.py

Reviewer Guide

Suggested review order:

1. mlx_audio/tests/test_server.py (especially reserved model_kwargs rejection case).
2. mlx_audio/server.py request parsing + validation flow.
3. mlx_audio/sts/voice_pipeline.py and mlx_audio/sts/tests/test_voice_pipeline.py.
4. mlx_audio/stt/models/qwen3_forced_aligner/__init__.py (scope-limited integration touchpoint).

[lucasnewman]

Closing for now. We'll reopen if work on #501 is resumed.