[clkao]

Add an optional Nemotron MLX adapter using mlx-audio's incremental mel frontend and cached Conformer encoder, with a small stateful RNNT decoder loop. WLK owns voice activity detection and stream boundaries; the copied encoder, second VAD and full-utterance second pass are removed.

Use the converted mlx-community/nemotron-3.5-asr-streaming-0.6b checkpoint and mlx-audio>=0.5.1,<0.6. The previous default points to a Transformers checkpoint and fails at startup with Model type nemotron3_5_asr not supported for stt. Sessions have independent caches and language prompts, while model loading and decoding stay on one dedicated thread. A shared mutex alone was insufficient: repeated sessions failed with There is no Stream(gpu, 2) in current thread. EOF flushes the final frontend/encoder frames once. Token times are encoder emission times, not forced word alignments.

Validation: The final head passes all required CI checks, including real-audio boundary scenarios; two MLX CPU scenarios cover partial-frame draining, pauses, EOF, interleaved sessions, thread ownership and language/context handling. The frozen extra installs mlx-audio 0.5.1 and completes real English/French/Chinese streams on M5. Ruff and lock checks passed; existing Metal, Torch and Transformers versions remain unchanged.

The completed M5 evaluation does not meet the agreed rule. All sixteen reports were measured on AC power (#444, benchmarks/evaluations/m5-20261107/). Head 1842f4c completes all 270 short-clip sessions and the ten-minute Chinese stream, with valid line timestamps and quality identical to September. EOF p95 is 71% to 94% lower than Whisper small and Qwen Metal in every pass. French passes against both baselines, but English WER is 19.27% versus 11.66% for Whisper small and Chinese CER is 22.62% versus 11.85% for Qwen Metal, beyond the one-point limit. The ten-minute stream has no repetition loop; its output contains one run of twenty Thai characters.

This PR remains stacked on #425; the research translation changes are already on main. The restored README, badge and published figures are unchanged.

[QuentinFuxa]

I measured this one together with #425 (series in benchmarks/evaluations/m5-20261107/) and ran it on real audio. It works well: correct text on the English and French recordings in audio_tests/, EOF delay between 0.02 and 0.17 s, valid timestamps, and three concurrent sessions run without the MLX thread error. In the FLEURS series it beats Whisper small MLX in French (15.0% vs 17.7% WER) and Chinese (22.6% vs 25.8% CER), with EOF p95 around 0.1 s.

I'd like to merge it. What it needs:

1. Make it independent of Add a native Qwen3 MLX streaming adapter #425. It currently includes the Qwen adapter, which is not ready (see my comment there). Could you rebase it alone on main, with only the Nemotron files?
2. Regenerate uv.lock on current main, it is the only conflict.
3. The output has double spaces between sentences (test. We want). Probably the token text joined with a leading space after punctuation.
4. English WER is 19.3% against 11.7% for Whisper small, so docs/backends.md should say it is a low-latency option that is strongest in French and Chinese rather than a general replacement.

One note for later: #450 pins mlx-audio<0.5 and this needs >=0.5.1, so the two cannot be installed together yet.

[clkao]

The double space is fixed. The decoder emits a space token, then a word token that starts with a space. The adapter now collapses each space run to one. In 60 cached FLEURS clips in English and Chinese, there are zero double spaces and the word error rate is the same.

docs/backends.md uses the m5 figures. French is better than both baselines. Chinese beats Whisper small. English is 7.6 points behind Whisper small.

[QuentinFuxa]

Merged, thanks. The space fix and the docs read well, and with the branch independent of #425 it went in cleanly. I re-ran the English and French recordings on the final branch: correct text, no double spaces, three concurrent sessions fine.