Threat-Preserving Representation: Why Agent Security Benchmarks Lie About Model Robustness