Accurate building instance segmentation in remote sensing imagery remains a challenging task due to boundary ambiguities and merging of adjacent structures in dense urban scenes. While convolutional and transformer-based networks achieve strong pixel-wise accuracies, they often fail to preserve boundary integrity and instance separability when adjacent rooftops are closely spaced or overlapping. To address this limitation, we propose a boundary-guided and merge-aware segmentation framework that enhances instance separation without sacrificing pixel-level precision. The framework introduces (1) a Separation-Preserving Loss for Instance Topology (SPLIT), a differentiable proxy for merge rate that penalizes overconfident building predictions along ground-truth boundaries, and (2) a lightweight, backbone-independent Contour Head that predicts continuous boundary maps to reinforce structural separation. We quantify topology-level errors and evaluate merge severity across datasets. Extensive experiments on three benchmark building segmentation datasets, Massachusetts, INRIA, and WHU, using multiple CNN and Transformer backbones (U-Net, SegFormer, and BuildFormer) demonstrate that integrating the proposed components significantly reduces merge errors while preserving instance-level accuracy and boundary quality. The proposed framework offers a simple yet effective approach for preserving object topology, providing practical and robust building instance segmentation in complex urban environments.