Retries can recover a temporary dependency failure, but they can also repeat a committed side effect. Marlow Supplies receives an approval event, updates its fulfillment system, and then loses a response. A new random operation identifier on the next attempt would make the same approval look like new work.
Keep stable business identity, outcome recording and broker acknowledgment as deliberate parts of the contract. Bound retries and concurrency when a dependency is unhealthy. Persistent invalid data needs correction or a recoverable failure path, not an endless immediate retry. These tasks combine supported subscription configuration with architecture decisions. The simulator does not execute distributed retries, real dead-letter delivery, exactly-once settings or external fulfillment calls. It checks the local declaration and the reasoning about where repeat processing remains possible.
Preserve business intent across retries and replay.