…tart fixes

Field reports, same session:
1. Callouts only played on the Live Race Control page — the whole audio stack
   (player, speech queue, start/end tones, lap detection) lived in that screen
   and died on navigation. It now lives in an app-shell controller
   (raceDayAudio.svelte.ts) mounted once from App.svelte, consuming the same
   app-wide live stream the global header uses: a running race is audible from
   marshaling, rounds, results — anywhere. The Live page keeps the Callouts
   toggle + control-click unlocks, driving the shared instance. Bonus: page
   navigation no longer remounts the detectors, so returning to Live mid-race
   can never re-buzz the start tone.
2. ~13s of race with no tones or callouts. MEASURED the pipeline first: a
   scratch heat instrumented at 200ms poll showed passes visible in the live
   projection 0.22-0.50s (median 0.33s) after the crossing — the data path is
   sub-half-second. The stall is the browser TTS engine: its first-ever
   utterance can take seconds or die SILENTLY while the backend spins up, and
   the serialized queue dammed every later callout behind the corpse. Two
   fixes: warmUp() speaks a volume-0 utterance on the RD's first gesture (the
   same one-time unlock that resumes the AudioContext), and a per-utterance
   dead-engine WATCHDOG (6s) that cancels + advances so one swallowed utterance
   can never silence the rest of the race.

Noted for a later slice: adapter-appended passes carry no recorded_at (the
latency measurement had to poll live instead of reading the log).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>