The prompt voice is rewritten on main's ranking: the best usable voice
for the language (no last resort, the same variety) that is not the
person's; with none, the tone plays instead. The prompt is tracked: the
person's words clear it and stop it, it cancels nothing but the prompt
before it, its utterance is kept so its end event is not lost, and it is
over after a time fitted to its words if that event never comes. Web
Speech is unlocked by main's unlock_in_tap; the tone's audio is resumed
only inside a real gesture, and again at each tone.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
3e7b399f73