[TSOFTP-afk]

Motivation

registry.py has preset a "stt" category since the 2026-03-31 modularization refactor, but it never received an implementation: zero plugins, zero create("stt", ...) call sites. Meanwhile the local SenseVoice/FunASR engine (#609, for #604) lives hard-coded inside server/asr_server.py, so the engine cannot be selected or replaced without editing the WebSocket handler.

This PR completes that design: STT becomes a registry plugin, and FunASR is the first registered engine.

What changed

• stt/base_stt.py (new) — BaseSTT: ensure_loaded() / transcribe(audio_bytes, sample_rate, is_final) / is_available().
• stt/funasr.py (new) — @register("stt", "funasr") FunASRSTT. The engine is moved verbatim from server/asr_server.py: lazy singleton, asyncio.Lock + threading.Lock pair, mutex-serialized inference, iic/SenseVoiceSmall + fsmn-vad, vad_kwargs={"max_single_segment_time": 30000}, trust_remote_code, <20 ms utterances skipped, rich_transcription_postprocess.
• server/asr_server.py — keeps the /api/asr WebSocket protocol byte-identical; the engine is now obtained through registry.create("stt", ...). Silent degradation is preserved: when funasr is not installed, the endpoint is simply not registered.
• config.py / config.yaml — new --asr key (default funasr). Also fixes the stale tts help text (cosyvoice / fishtts / indextts2 were removed from the codebase but were still listed in the help strings).
• registry.py — the unused "output" preset is corrected to "streamout" (the category output plugins actually register under).
• tests/test_asr_server.py — the two behavior tests (lazy singleton constructed once under concurrency; shared-model inference serialized) are migrated to the plugin, and a new protocol contract test locks the /api/asr message flow: config JSON → binary PCM16 frames → {"is_speaking": false} → reply {"text", "mode", "is_final", "timestamp"}. Tests remain dependency-free (fake modules injected into sys.modules).
• Docs — local ASR is now documented in README.md / README-EN.md / docs/api.md, with an explicit note that server/asr_server.py (user speech recognition) and avatars/audio_features/base_asr.py (lip-sync audio features) are unrelated despite the similar names.

Not changed

/api/asr wire protocol, the web/asr/ frontend, the avatar/TTS/LLM pipelines, and avatars/audio_features/* (lip-sync feature extractors).

Verification

$ python tests/test_asr_server.py
...
Ran 3 tests in 0.311s
OK

Engine latency, measured to document the feature (the engine itself is untouched; same environment for all runs — Windows, RTX 3060 Laptop 6 GB, torch 2.13.0+cu130, funasr 1.4.16, 8.3 s utterance @ 16 kHz, 7 timed runs after one warm-up):

session	Avg	Min / Max	Std	RTF
run 1	245 ms	236 / 254 ms	7 ms	0.029
run 2	131 ms	123 / 140 ms	6 ms	0.016
run 3	141 ms	120 / 164 ms	14 ms	0.017

The run-to-run spread is environment noise (GPU clocks/thermals), not a code difference — the engine code is identical before and after the refactor. The protocol contract test is the behavior-preservation evidence.

Refs #601, #604