Retry With Exponential Backoff and Jitter

Every model provider rate-limits you, and every naive retry loop makes it worse.

Implement compute_delays(attempts, base, cap, jitter_fn) returning the delay before each retry:

• Delay for attempt i (0-based) is min(cap, base 2*i), then jittered
• Jitter is full jitter: the actual delay is jitter_fn(0, d) where d is the capped exponential delay
• Return a list of attempts delays

jitter_fn(low, high) is injected so this is testable: in production it would be random.uniform.

Why full jitter and not "add a bit of randomness": if a thousand clients are rate-limited by the same overloaded service at the same moment, a deterministic backoff means all thousand retry at exactly the same instant, and again at the next, and again. Full jitter spreads them across the whole interval, which is what actually lets the service recover. Adding ±10% does not.