Leave-one-cohort-out survival-signature benchmarking with built-in leakage audits (fully nested re-selection, permutation calibration, size-matched random-panel nulls), extracted from a nine-cohort breast-cancer gene-signature study.