Every agentic assessment of this requirement under one review, newest first.

Both the production run and the playground re-tests, because comparing the two is the tuning question — is_test tells them apart. Sorted on created rather than started_at, which is null until a dispatched run is picked up.

review_id is required rather than derived from the requirement. A packet reaches its review through evidence_review_tasks, which carries no uniqueness on evidence_packet_id — so "one review per requirement" is convention, and a second task would silently widen this history to runs the caller did not open. Naming the review keeps the answer to the one asked about.