Choosing What to Observe: Budget-Paired Counterfactual Audits for Credible Evaluation of Visual Agents — ThinkLLM