Collaborative Knowledge Graphs (KGs) rely on evolving, “soft” constraints, making automated repair a challenge for rigid symbolic methods. In this paper, we propose a structure-aware neural framework that learns to repair violations directly from historical edit patterns. We perform a systematic study of graph encoding strategies, specifically comparing flattened predicate-as-node representations against multi-relational graphs equipped with a custom dual-representation module. We benchmark these encodings across different GNN backbones and feature initializations. Our experiments show that the Multi-Relational GIN yields the most robust performance, surpassing strong symbolic baselines by over 29% in historical fidelity and achieving a 93% validity rate. Furthermore, ablation studies indicate that topological features, such as role embeddings and edge attributes, are significant performance drivers, often outweighing semantic text features. These findings suggest that effective repair depends heavily on precise topological modeling, reinforcing the premise that structure is the signal.