Feature flag cache staleness after kill-switch

An incident is in progress: a buggy 'new_comment_ranking' code path is throwing for 4% of comment loads. You flip its kill-switch flag to OFF at 21:10. The error rate drops for some pods but ~30% of pods keep throwing the same error for another 8 minutes, then they too recover with no further action. The flag SDK caches flag values locally and polls the flag service every 5 minutes; the flag service confirms OFF immediately. Triage and explain what happened and how to prevent it.