Detects when a user message corrects a prior one, builds an active state where corrections override originals, and injects only the live instruction set into agent context. Prevents model drift when users change their minds mid-session.
