Schema Is the Benchmark: Testing Cross-Schema Transfer in Graph Foundation Models

Graph foundation models promise to pretrain once and adapt across graph tasks, domains, and feature spaces, but many evaluations publish a single dataset-specific rendering, making it hard to separate graph transfer from fit to that rendering. We argue that cross-schema transfer is a missing test: can a model preserve task behavior when labels, evidence, and splits are fixed but the graph grammar changes? SchemaWeaver is a benchmark protocol, not a new model; its goal is to make representation-level transfer measurable before leaderboards conflate schema engineering with model capability. It turns relational databases and text-attributed graphs into semantically matched schema views, then requires task contracts, executable leakage audits, adaptation budgets, and metrics for utility, robustness, and calibration. A reproducible synthetic pilot illustrates the measurement issue: under a deliberately schema-local extractor, moving predictive attributes away from the prediction object induces a measurable AUROC range, while an oracle contract control changes the interpretation by spending an explicit adapter budget. The paper gives a concrete path from fixed-schema graph benchmarks to a community stress test for genuinely transferable GFMs.