Multi-model registry and adapter layer for realtime-0.5b, tts-1.5b, tts-7b

[Copilot]

Refactors the hardcoded single-model runner into a multi-backend TTS system. The model field in OpenAI-compatible requests is now resolved via a registry instead of ignored. Longform models (1.5B/7B) are registered with API plumbing but fail gracefully with 501 until a real backend is wired in. Existing realtime-0.5b behavior is fully preserved.

New runner/ package

• model_registry.py — ModelProfile dataclass, 3 model profiles, alias map (tts-1 → realtime-0.5b, etc.), resolve_model_key(), get_model_profile()
• adapters/base.py — EngineAdapter ABC with is_available(), capabilities(), synthesize(), stream(), health()
• adapters/realtime_demo.py — RealtimeDemoAdapter wrapping existing subprocess demo. Extracted detect_device(), apply_overrides(), build_realtime_demo_cmd() from run_realtime_demo.py
• adapters/longform_native.py — LongformNativeAdapter scaffold. is_available() returns False; synthesize() raises BackendUnavailableError. No fake streaming
• adapter_factory.py — make_adapter(model_key) dispatches by family/loader_mode
• types.py — SpeechRequest/SpeakerTurn pydantic models with per-family validation
• errors.py — UnknownModelError, CapabilityError, BackendUnavailableError, InvalidRequestForModelError

Modified files

• overrides/app.py — /v1/audio/speech resolves model, returns 400 (invalid combo), 404 (unknown model), 501 (missing backend). /stream rejects non-realtime models. /config and /health expose model registry info
• scripts/download_model.py — Accepts --model with registry resolution. Default unchanged
• scripts/run_realtime_demo.py — Now a thin shim delegating to runner.adapters.realtime_demo
• scripts/run_server.py — New generic launcher: --model realtime-0.5b|tts-1.5b|tts-7b

Backward compatibility

All existing entry points work unchanged:

uv run python scripts/download_model.py           # still downloads realtime-0.5b
uv run python scripts/run_realtime_demo.py --port 8000  # still launches realtime demo
curl -d '{"model":"tts-1","input":"Hello"}' /v1/audio/speech  # still maps to realtime-0.5b

Remaining TODOs

• longform_native.py:_load_backend() — replace placeholder ImportError with real backend import
• longform_native.py:synthesize() — wire actual inference
• run_server.py — add longform serving path

Tests

46 new tests covering registry, alias resolution, per-family validation, adapter factory, graceful degradation, and error classes. All run without GPU/model weights.

---

📱 Kick off Copilot coding agent tasks wherever you are with GitHub Mobile, available on iOS and Android.