The strongest scientific case is architectural complementarity. Neural models learn distributed statistical regularities; graphs preserve explicit entities, typed relations, provenance, and compositional paths. Embeddings offer scalable approximation, while symbolic rules and graph traversal expose interpretable constraints. Yet the combination is not automatically superior: graph incompleteness, noisy extraction, relation imbalance, leakage during evaluation, and distribution shift can produce confident but invalid inferences. The central research question is when relational inductive bias improves generalisation enough to justify added system complexity.