Recent advances in 3D reconstruction have enabled the automated generation of city models from LiDAR point clouds and aerial imagery. Although these methods produce visually convincing models, they are primarily designed for visualization purposes and often lack the geometric and topological quality required for Urban Digital Twin applications, such as physics-based simulation, semantic analysis and predictive modelling. Automatically reconstructed LoD2 building models frequently exhibit defects including open boundaries, non-manifold configurations, inconsistent face orientations and degenerate geometries. This work presents a graph-guided framework for the topological repair of LoD2 building models. Mesh connectivity is represented as a graph, allowing boundary, adjacency and local geometric information to identify defective regions and guide robust repair operations. By combining topology-aware reasoning with geometry-constrained graph processing, the proposed method restores watertightness, improves 2-manifold consistency and preserves the structural integrity of complex building surfaces. Unlike conventional geometry-based post-processing techniques, the proposed approach explicitly targets topological correctness, producing high-quality building models suitable for simulation and analysis within Urban Digital Twins.