Reproduce tool failures and check the side effects that a successful response can hide. A lost response can turn a retry into a duplicate shipment. The browser lab compares scripted policies, not LLM performance. No real shipments are made.