A retryable error's agent_end fires with the failed assistant message
(stopReason === 'error') the instant #handleRetryableError schedules a
retry (auto_retry_start), before the retry has a chance to recover.
sendErrorNotification read that transient agent_end the same as a real
final settle, so error.notify=on raised a 'Stopped with error' toast
even for turns that went on to succeed on retry.
Track a #retryPending flag set on auto_retry_start and consumed
(check-then-clear) by sendErrorNotification, so exactly one mid-retry
agent_end is suppressed per attempt. #handleAutoRetryEnd also clears it
directly on both success and failure so a recovered retry never leaves
it stuck; consuming it on read (rather than only via auto_retry_end)
keeps it self-healing for the classifier-refusal short-circuit path,
which can return from #handleRetryableError without ever emitting a
fresh auto_retry_end.
Mathews-Tom
2a3b6856ff