v0.7.0: "Hey Vox" trigger, mid-speech command detection, earlier TTS loading

[JRufer]

Summary

Release merge for v0.7.0. Adds a "Hey Vox" voice-command trigger, recognises commands while the user is still speaking, and loads TTS models earlier. It's a minor bump because it adds a trigger keyword, a setting and new mid-speech behaviour.

PR	Summary	Detail
#147	"Hey Vox" trigger keyword	Works alongside "VoxCtrl" / "Vox Control", and also accepts near misses like "hey box" and "hay vox". Triggers now need word boundaries. "… control" only triggers at the start of a dictation when the first word sounds like "vox", so "Take control, say hello" is no longer sent to a Say target. Matching target names in text with characters like İ or Ü can no longer crash the delivery thread.
#148	Voice commands recognised mid-speech	Early passes transcribe the first 6s of a recording, so a command's overlay appears and its TTS model starts loading before the hotkey is released. The final transcript alone decides routing. If it doesn't confirm an early command, the overlay is withdrawn. Early passes skip the hotkey's OpenAI rewrite, and are skipped entirely for the remote backend and CPU-only medium/large Whisper models. New setting: features.early_command_detection, default on.
#149	No S1-mini rewrite for spoken text	Text going to a Speak target is read aloud word for word, and the first word comes sooner.
#150	TTS loads earlier, without loading at every launch	Preloading now also works in always-loaded mode. The per-engine prewarm setting still decides whether the model loads at startup. Switching to always-loaded with prewarm on loads the model right away.
#151	Overlay never overlaps	The target visualizer is hidden while SYSTEM RESPONDING shows. Recording takes priority over SYSTEM RESPONDING. The visualizer's fade-out still plays, and the pills stay inside the overlay window.
#152	Test fix	A clipboard test ran the real wl-copy: it replaced the developer's clipboard and made piped cargo test hang.
#153	Version bump	Updates Cargo.toml, Cargo.lock, package.json, package-lock.json and tauri.conf.json.

Note on 1f13d62

The history includes 1f13d62 and its revert, 29ab1f8. That commit bundled all of the above into one change and had problems the split PRs fix:

• a billed OpenAI request on every interim pass;
• TTS models loaded at every startup, making the prewarm toggles do nothing;
• on-demand TTS memory mode effectively disabled;
• an early-match fallback that could send ordinary dictation to a command target.

Together, the commit and its revert change nothing; the PRs above are what ships.

Test plan

• cargo test for voxctrl-app, voxctrl-config, voxctrl-routing, voxctrl-tts and voxctrl-inference on the merged development: 426 pass
• npx vitest run: 244/244 pass
• CI (Linux + Windows cargo check/cargo test, frontend check) green on every PR above, re-run after each stacked PR was retargeted to development
• Version reads 0.7.0 across all five version files
• Release build with default features (moonshine/parakeet/inflect-micro); not run yet
• On a real desktop:
  • "Hey Vox, say hello" shows the command overlay before the hotkey is released and is spoken verbatim;
  • dictation without a trigger is typed as before;
  • with on-demand TTS, the model does not load for ordinary dictation.

Tagging v0.7.0 is left to you, as with previous releases.

🤖 Generated with Claude Code