The point about reacting differently to 429, 5xx and timeouts is the part many dispatchers skip. One suggestion on the per-endpoint breaker: when it half-opens, it may be worth sending probe traffic from a small dedicated lane rather than the main queue, so a recovering consumer isn't hit by a backlog burst the moment it closes. How would you handle ordering for consumers that need per-entity ordering while retries are being delayed by the breaker? Also curious whether Retry-After values are capped, since a misconfigured consumer could otherwise park deliveries for hours.