Three checks before I merge code an AI agent wrote

An agent reports done, tests green, PR open. Three concrete checks catch the failures that unit tests and the agent's own self-review routinely miss.