판 이력 — Labels Are Not Endpoints: Treatment Leakage and Construct Validity in MCP Agent Security Evaluation | AIChainDay