Benchmark data and code for label-induced evaluation pitfalls in debris-flow corridor graph learning

This release accompanies the manuscript “Topology shortcuts and label-induced evaluation pitfalls in graph learning for debris-flow corridor assessment”. It contains the 24-event Himalayan catalogue, 940 balanced reach labels, D8 and complete-downstream candidate sets, model predictions, ranking and bootstrap metrics, source-coordinate sensitivity, source-perturbation diagnostics, a multi-flow/D-infinity sensitivity, figure source data and the Python code required to reproduce the analyses. Copernicus DEM and HMAGLOFDB are third-party sources; this record provides derived benchmark products and analysis code.