Voice Transcriber v0.6.0 — Choice

Dictate a thought, review it, and paste it anywhere on Linux—with the active transcription boundary visible.\n\nWhat changed:\n- A typed provider contract now normalizes capabilities, cancellation, limits, errors, languages, and privacy boundaries.\n- Groq remains the supported Flatpak path with a user-managed key.\n- A source-only whisper.cpp prototype is available behind an explicit feature flag; it is disabled in Flatpak and is not claimed as supported or universally offline.\n- The first run, Settings, desk, diagnostics, and config checks all expose the selected provider boundary.\n- The reproducible LibriSpeech test-clean benchmark receipt reports 37 word errors over 627 reference words (5.90% WER), p50 1043.810 ms and p95 1508.576 ms on the named GitHub runner.\n\nVerification:\n- Exact-tag CI run 33016125424: Ruff, compilation, 50 deterministic tests.\n- Exact-tag Flatpak run 33016125188: build, install, version/help/doctor, GTK onboarding and accessibility smoke, uninstall.\n- CodeQL run 33015923746 passed Python, Actions, and JavaScript with no open alerts.\n- Bundle SHA-256: fd4e59b2af2f72f9158ce8275fde30fddb3ff90b9fb7d7888770d53e9874e3ea.\n\nKnown limits:\n- x86_64 Flatpak only. Real microphone, PipeWire/PulseAudio, X11/Wayland, and tester-owned Groq-key acceptance gates still need physical Linux reports.\n- The local prototype is source-only and depends on user-supplied whisper-cli/model files.