Epistemic Humility in Agent Evals: What Happens When Retrieved Evidence Contradicts the Model's Prior Beliefs