Count start failures in an independent streak inside scheduleRestart so
config failures, a missing server entry, and crashes after a bounded
preflight fallback reach the notification threshold. Bounded preflight
retries opt out because they surface their own dialog. The callback runs
on a forked fiber so an open dialog never holds the instance mutex, and
the pool shows an async message box whose Copy Diagnostics button copies
the failure reason and backend log path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>