Streaming transcription, speaker labels and endpointing from open weights on a CPU

The models were the easy part and they are swappable. The problem is that the three signals arrive on three different latencies, and a closed model that emits them as tokens in one vocabulary gets the ordering for free. So the project is the merge. The rule that makes it usable is that committed output is append only: once a commit is emitted nothing later contradicts it. A speaker label cannot exist before enough audio has been heard, so early words commit with speaker: null and a later event fills them in retroactively. Committing a guess there would be unfixable; leaving a gap is not. A consumer that ignores every partial still gets a correct transcript, which is what keeps simple integrations simple.

And installable as hearwrite[onnx,turn,server]. Measured on an M4, CPU only: emission delay p50 0.28s to 0.52s, false endpoint on 5.0% of mid-thought pauses, 6.8% and 4.4% WER on LibriSpeech dev-clean, all reproducible from docs/evaluation.md. Diarization is on the front page as experimental, because three people on one laptop microphone finds two speakers and splits turns mid sentence. Still active.