Voice input: on-device STT default, API-key fallback, keybinding + level animation

Thinking blocks expand by mouse-click only (ReasoningAction::Expand is an InteractionTarget; no keyboard path). Voice input should follow the same inline pattern: on-device speech recognition where the OS offers it, with a wired STT API-key fallback (add the key slot), a single keyboard command to start/stop, and a live level/listening animation in the composer while sound is accepted. Keep the collapsed transcript rendering untouched.

Start by reading the ReasoningAction::Expand and InteractionTarget paths, then trace the composer and keyboard-command handling. Identify where OS speech recognition, the STT API-key slot, and listening-level animation belong. Done means start/stop works from one command, on-device recognition is preferred with fallback available, the animation appears while accepting sound, and collapsed transcript rendering is unchanged.

Written by the indexing model from the issue text.