The Judge Reads Tokens: An Adversarial Evaluation of an Eval Toolkit's Own Scorers