Retry loops fixed for spent one-time tokens

The core run-engine pipeline is getting smarter about when to give up. Previously, if a job failed because a one-time token had already been consumed, the engine mistook the database collision for a standard idempotency key conflict. This triggered a pointless retry loop, wasting compute cycles on a task that was permanently doomed.

By explicitly checking which database field caused the collision, the engine now instantly rejects reused one-time tokens with a specific error, while continuing to safely retry true idempotency conflicts. This ensures cleaner error handling and stops the system from fighting a losing battle against exhausted credentials.