v0.2.0

Release v0.2.0

Notable changes since v0.1.2:

CLI

• aureka download: pre-fetch all model weights (Kokoro + faster-whisper)
• aureka benchmark: measure ASR/TTS/LLM speed and write a Markdown report
• aureka type --no-streaming: opt out of streaming
• aureka ui (settings UI via pywebview), aureka type/speak --speed

ASR

• VAD-segmented streaming via silero-vad: partial transcripts arrive while
  you are still talking
• In refine/translate modes, partials are stderr-only and only the final
  refined text is injected into the cursor
• Phase events (transcribing / finalizing / refining) so the client can
  show what the daemon is doing
• faster-whisper model is now configurable via [asr] model in config.toml

TTS / Daemon

• POST /speak endpoint: shared warm Kokoro pipeline across clients
• POST /reload: re-read config without restarting the daemon
• aureka speak is now daemon-aware and streams over HTTP when the daemon
  is up
• System tray menu (pystray) with daemon controls
• Cross-platform launch-at-login (macOS launchd, Windows Task Scheduler)

LLM refine

• Stricter system prompt + few-shot examples to suppress reasoning prelude
• Detect truncated-mid-thinking responses and fall back to raw transcript
• max_tokens / thinking_budget knobs in [llm] config

BREAKING

• TheWhisper backend removed (the PyPI package was a placeholder, the
  code path never actually loaded). [asr-thewhisper] extra dropped.
• Default ASR model changed from large-v3 to medium. Override with
  [asr] model = "large-v3" if you want the old precision.
• aureka.models.MODEL_REGISTRY (dict) replaced by model_registry() (fn).
• aureka.device.resolve_asr_backend removed.