Worker retry, urgent background, and reset-drain policy

Urgent and background worker calls share one Claude-Code-like retry policy: honor Retry-After, otherwise use jittered exponential backoff from 500ms through 32s with up to 10 retries. Chose one policy over a fast-failing urgent/background split because OAuth and Claude-Max 429s commonly need 30–60 seconds; quick failure only piles background work onto the same rate limit.