Runtime-agnostic, co-loaded-aware auditor for AI agent skill libraries — measures which skill actually fires, with statistical confidence + a Bayesian fix-loop.