How can generated state support safe routing actions when records change concurrently?

An executable package can define an assignment command with `request_id`, `destination_queue_id`, `expected_version`, `reason_code` and `idempotency_key`. The request record can hold `current_queue_id`, `status` and `record_version`. Queue references should resolve to the same identifier space across all fixtures.

The simulator can apply an assignment only when the supplied version matches current state. A successful command updates ownership, increments the version and appends an event containing `event_id`, `request_id`, `previous_queue_id`, `new_queue_id` and `caused_by_action_id`. Those changes should commit together.

If another actor has already updated the request, the tool can return a conflict with the current version. Repeating a successful command with the same idempotency key should return its recorded result without producing another transfer.

MirrorFirm could generate initial snapshots and scripted concurrent events in JSONL, with larger state tables in Parquet where appropriate. Builders can replay a dispatcher’s reassignment between an agent’s read and write, then inspect whether the agent refreshes its evidence. This makes concurrency a defined part of the task rather than an incidental simulator error.