Defends a supervised pipeline against four leakage classes ordinary train-test-split audits miss: (1) LOFO sweep that surfaces a feature whose removal collapses test performance (target-encoding / post-event signature); (2) hub-firewall isolation for crosswalk / multi-source data where one "hub" source connects others and its representation leaks to held-out sources; (3) group-aware splitting (GroupKFold, GroupShuffleSplit, leave-one-group-out) so multi-row-per-entity data (patient visits, user sessions, customer orders) does not put the same entity in train and test; (4) no-row-in-two-splits invariant via row-hash check after split. Triggers when fitting on multi-source …
