Traffic STGNNs across Sensor, City, and Time Shifts: Routing Concentration Tracks Sensitivity to Sensor Dropout

Traffic forecasting benchmarks measure accuracy within a single city and test distribution, leaving open how architectures behave when sensors fail, when a model is transferred to another city, or when weekday and weekend traffic differ. We evaluate eight STGNN architectures under these three shifts using frozen checkpoints; for transfer across cities, target graph information is withheld so that fixed and adaptive graph models have the same structural access. We find that fragility under sensor dropout follows routing concentration rather than the use of adaptive adjacency itself. Across four adaptive backbones, mean row maximum routing mass correlates with dropout degradation at Pearson $\rho=+0.979$, and the relationship replicates on a second source city. Graph WaveNet lies at the concentrated end of this spectrum, while AGCRN and STAEformer stay close to DCRNN when sensors are zeroed. The other shifts separate from this pattern: MegaCRN performs best when transferred to another city, yet MegaCRN and STAEformer show the largest weekday and weekend performance gaps. In this backbone bank, strong performance under one shift does not imply robustness under another. We release a registry of frozen checkpoints, evaluation code, and a decision table mapping each backbone to its main failure mode.