A machine-unlearning audit project that tests whether recovery-radius evaluations measure residual memory or quirks of the evaluator. The current paper is a scoped evaluator-invariance case study with a native-verified candidate portfolio as the repair.