Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic