Constraint Memory and Long-Horizon Failure in Artificial Intelligence: Why Alignment Decays Even When Local Behavior Succeeds | Synapse