Repeated cross-fitting finds one split-stable benchmark sensitivity amid heterogeneous dataset effects

This technical preprint reports a prespecified repeated cross-fitting audit of protected-feature construction in a six-dataset public microbiome robustness benchmark. Across 20 outer-split seeds, one WT2D random-forest zero-inflation condition showed a stable material global-minus-fully-nested AUROC difference, while the remaining dataset effects were heterogeneous, small, or null. The release includes 20,700 out-of-fold predictions, a locked protocol, tests, independent verification, checksums, the manuscript, and an executable reproducibility package. It does not establish biological mechanism, clinical validity, classifier-training leakage, source-author error, or invalidity of the source paper.