Graph Tokens before Graph Propagation: Atom-Keyed Retrieval Interfaces for Graph Foundation Models

Graph foundation models for language-facing applications must decide not only how to propagate over a graph, but also what object should become a graph token and how that token should connect to the text an LLM reads. We study this interface in long single-document QA. Our position is that the unit a graph retriever \emph{ranks} and the unit a reader \emph{consumes} should be decoupled by a deterministic source map. Atom-Keyed Chunk Retrieval (AKCR) treats atomic Davidsonian propositions as graph nodes, uses typed edge hints such as entity overlap, scene index, and temporal anchors as schema primitives, ranks atoms by query similarity, and maps selected atoms back to source chunks as the LLM payload. On full-book NarrativeQA (40 books, 1{,}169 questions), AKCR gives a bounded but useful efficiency result. With the same GLM-4.7 reader, AKCR at 24K characters scores $0.5543$, above a chunk-cosine retriever at the same budget ($0.5227$), while remaining below the full-book oracle ($0.6450$). With GLM-5.1, retrieval exceeds the degraded full-context oracle, but we read this as long-context degradation rather than as evidence that retrieval beats context. A scoring-rule ablation favors atom-best max pooling, while two negative results set the boundary of the claim: a learned typed-edge propagator drives $\mathrm{ppr}_\alpha\!\to\!0.95$, effectively disabling propagation in this single-book pool, and feeding atoms directly to the reader plateaus near $0.46$. We therefore argue for \emph{graph-token first} evaluation protocols for GFM--LLM interfaces: test the token, source-map, and payload design before attributing gains to propagation.