Interface drift is a confound in graph-RAG evaluation in which an apparent gain from a graph-aware retriever is actually driven by simultaneously changing the retrieval interface — the encoder, dense candidate pool, cutoff $k$, matching rule, or split policy — rather than by graph traversal itself. When the graph policy and the interface change together across compared systems, traversal effects cannot be isolated, so reported improvements may reflect a richer interface rather than graph reasoning. The strict-parity contract is the corrective: fix all interface components across systems so the only varying factor is the graph component.