[TheGreatAxios]

What changed

Click the mic, then click stop before answering the browser's permission prompt. Nothing is listening yet, so stop does nothing, and when the prompt is allowed the page starts listening anyway: the button reads "Stop dictating", a stream is held and recognition runs. Opening a second composer in that window gives two live sessions.

• The dictation session now exists from the moment the mic is asked for, not from when it is granted.
• Stop (or leaving the page) during the prompt marks that session cancelled.
• When permission arrives on a cancelled session, the tracks are ended and recognition never starts.

This changes only what the page does. Safari keeps the device indicator on for a while after a recognizer ends, on main and here alike; that is the browser, not this code.

How it was verified

• lint, format:check, typecheck, pins:verify pass; bun run test 328 pass. The new tests fail on main.
• Ran main and this branch against a local hub with a fake model, driving the mic with a scripted permission prompt. On main the track stays live and the button reads "Stop dictating" after allowing; here the track is ended and the composer is idle. Same for navigating away mid-prompt and for a second composer.
• Reproduced by hand in Safari on main; normal start and stop cycles on this branch traced clean (recognizer ended, tracks stopped, context closed).

[corbits-triage]

Triage: Pass

No changes needed from you. A maintainer will review this pull request.

Changed since the last verdict (Unsure at 180112f): 1 new commit.

[corbits-triage]

No defects reported.

[corbits-triage]

No defects reported.