Use redacted cursor neighborhood and pause-aware chunks for more natural polish,
validate protected terms with retry/local fallback, and structure bilingual prompts
for consistency and provider prefix caching.
Harden polish so question drafts stay questions, add local density
routing with style-specific degrade, expand fun style packs, and add
end-to-end FlowTrace logging plus offline guard eval scripts.
- Add FlowUtteranceEndCoordinator with 350ms silence drain and 150ms post-roll
- Extend FinalChunkRecovery for short/empty final chunks in chunked pipeline
- Snapshot partial at mic stop and guard final transcript in FlowSessionManager
- Unify tail drain presets (iosFlow/macMLX) and expand diagnostics
Co-authored-by: Rocky <hkgood@users.noreply.github.com>
Reduce perceived latency from key release to final text:
- Adaptive chunking: 2.5s first chunk + 5s follow-ups so short
utterances start on-device recognition while still recording.
- Session-level ASR warmup and audio-format cache reuse to remove
per-utterance cold-start of SpeechAnalyzer.
- Mirror live pipelined partials to the keyboard transcript line via
a new flow.transcriptionPartial App Group key + Darwin ping.
Also commits the accumulated custom language model, Flow session,
keyboard extension restructure, and Xiaomi MiMo provider work in
progress on this branch.
Replace MLX GPU inference with CoreML bundles so transcription continues
while the host app is backgrounded. Adds model download and warm-up,
vendored Qwen3Speech, and updates onboarding, settings, and copy for the
~1.6 GB CoreML package (iOS 18+).
Rocky