feat: add Apple Silicon Qwen3-TTS MLX backends

[groxaxo]

Summary

Adds a native Apple Silicon serving path while preserving the existing VibeVoice CUDA/realtime workflow.

• serves the exact aufklarer/Qwen3-TTS-12Hz-1.7B-Base-MLX-8bit bundle through Speech Swift
• serves mlx-community/Qwen3-TTS-12Hz-1.7B-Base-4bit through the model-compatible mlx-audio==0.3.0 loader
• adds an adapter-driven FastAPI application for native/non-realtime models
• adds an arm64 macOS bootstrap, model aliases, reference-audio conditioning, health/config/model endpoints, and OpenAI-style speech output
• keeps the vendored realtime VibeVoice application isolated so it cannot advertise or route requests to a Qwen adapter it never loaded

Inconsistencies fixed

1. macOS was documented, but the active registry and launcher only supported PyTorch VibeVoice profiles.
2. scripts/run_server.py exited instead of serving non-realtime adapters.
3. Model sources were coerced to local paths, breaking Hugging Face ID fallback on a fresh Mac.
4. Registry/launcher imports eagerly pulled PyTorch backends into a Qwen-only Mac process.
5. The old realtime app could resolve model names while still synthesizing with one fixed realtime engine.
6. An empty/uninitialized VibeVoice submodule could pass the old bootstrap directory check.
7. The VibeVoice and MLX dependency stacks require incompatible Transformers and Hugging Face Hub generations; they are now isolated as mutually exclusive uv extras.
8. mlx-audio expects reference audio as a loaded/resampled MLX waveform, not a filesystem string; the adapter now performs that conversion explicitly.

Architecture

• realtime-0.5b: unchanged subprocess-based VibeVoice demo, including its web UI and streaming route.
• qwen3-tts-mlx-8bit: Speech Swift CLI, using the exact requested model ID.
• qwen3-tts-mlx-4bit: in-process mlx-audio adapter with lazy model loading and serialized MLX generation.
• one active model per process; a request selecting another model returns HTTP 409 instead of silently using the wrong backend.

Local validation

• python -m pytest -q test_macos_mlx.py — 17 passed
• python -m compileall -q runner scripts main.py test_macos_mlx.py
• TOML parsing and dependency-stack assertions
• AST parsing and duplicate top-level definition scan across the patched Python files
• bash -n scripts/bootstrap_uv.sh scripts/bootstrap_macos.sh
• compare check: one commit ahead of main, exactly 15 intended files, no accidental backend stubs

No GitHub Actions workflow was added or used.

Remaining target-hardware check

A real model download and Metal/Speech Swift synthesis smoke test must be run on native Apple Silicon macOS; the implementation environment used for this patch is Linux and cannot execute MLX/Metal or Speech Swift.