Really solid approach, the dependency-graph angle catches a whole class of leakage that column-correlation checks miss entirely. One kind it won't catch though: row-level leakage, like the same census tract showing up in both the train and test fold across different years of a longitudinal file. No column is derived from another there, so the lineage manifest looks clean, but the model is still seeing the answer wearing a different row index. Worth a companion check that groups by entity ID before the K-fold split runs, not just the column graph.