Known-Shortcut Interventions under Mixed Graph Distribution Shifts: A Diagnostic Study

Graph classifiers can exploit spurious correlations in both background topology and node attributes. We study how knowledge of a shift-prone feature block affects robustness under such distribution shifts. We introduce RedHouse, a synthetic graph-classification benchmark with independently controllable correlations between the label and the background graph family or node colors. Under a protocol with 300 labeled training graphs, targeted color-block randomization substantially improves worst-environment accuracy over the evaluated baselines. A prior-matched comparison shows that a simple GIN with the same finite-warmup intervention (T_w=100) achieves 95.7±2.6% worst-environment accuracy across ten seeds, compared with 93.6±4.2% for the full MixedStable framework; these results do not establish an advantage of the full framework. We distinguish persistent randomization from finite warmup followed by shortcut release: the main low-label setting primarily evaluates sustained shortcut suppression, whereas longer post-release training reveals substantial performance decay and seed-dependent failures. Additional experiments on external benchmark variants show heterogeneous outcomes, including failures of the full framework. Our findings identify targeted intervention on a known nuisance block as the main source of the observed RedHouse gains and motivate evaluations that match prior information, training budgets, and checkpoint selection. They do not establish that a short warmup reliably yields lasting invariance or that mixed shifts necessarily cause stronger shortcut dependence than feature shifts alone.