Occlusions challenge the scan-to-BIM pipeline; while point cloud completion methods show impressive results on synthetic benchmarks, performance on real-world scans lags significantly. This paper investigates if this is caused by bounding box information leakage during input normalisation, letting models exploit implicit context unavailable in practice. This leakage is rectified by normalising partial inputs to create an unbounded completion task. Modifications are introduced to aid the model in learning object bounds and point confidences. Experiments demonstrate existing models rely heavily on leaked bounds, whereas the proposed adjustments significantly improve performance in realistic settings, achieving a 25% and 40% reduction in Earth Mover’s Distance and Chamfer Distance. Findings establish that addressing boundary leakage is critical for real-world applicability, providing scan-to-BIM practitioners with methods that generalise across domains and enhance downstream tasks such as meshing. This paper establishes that future research must focus on completion methods that operate without implicit spatial priors.