YazSes v2.1.0-dev.3 — Wave D hardening + hard-tier seams (developer preview)¶

Pre-release · not published (publish workflows are pinned to v0.x/v1.x; v2 dev tags do not ship). v1.4.1 remains the stable release. Every feature is OFF by default.

Consolidates v2.1.0-dev.1/2 with a hardening pass and two pure safety/decision seams for hard-tier features whose heavy engines stay deferred. Design in design/adr/adr-v2-021..022.

New (all off by default / internal)¶

• Atypical-speech LoRA held-out gate (ADR-v2-021, [personalize] lora_min_improvement) — a pure gate that applies a corpus-trained LoRA adapter only when it beats the un-adapted baseline on held-out WER by a set margin. Reuses the ADR-014 held-out discipline so a bad adapter can never silently degrade recognition. The fine-tune itself stays deferred (heavy); the safety spine ships now.
• Codec Streaming seam (ADR-v2-022, [codec]) — a pure select_engine that routes to a streaming neural-codec engine (Kyutai/Mimi) only when explicitly enabled with a real backend, otherwise keeps faster-whisper. The codec engine is lazy behind the codec extra; the default multilingual path is unchanged.

Hardening¶

Every v2 pure core is now 100% covered: raised agent, pilot, polyglot, context (and the new adapter_gate, codec/select) to 100% with targeted edge-case tests; the one genuinely-unreachable defensive branch (polyglot mer-gate) is marked no cover with a proof note.

yazses features now lists 35 capabilities (new: codec).

Remaining Wave D backlog (needs real hardware/models)¶

Silent-speech (sEMG), pure-vision VLM screen commanding. These require a device / on-device VLM to build and verify safely, so only their design (research doc + ADRs) is captured.

Quality¶

947 tests green across Linux × macOS × Windows on Python 3.11 and 3.12. Heavy backends stay lazy behind optional extras; the default install and the v1 dictation path are unchanged.