Semantic segmentation of large-scale airborne point clouds traditionally relies on labor-intensive 3D manual annotations. While recent zero-shot methods attempt to alleviate this burden by distilling knowledge from 2D Vision–Language Models (VLMs) via 2D-to-3D projection, they suffer from performance degradation in complex urban environments. Specifically, lacking 3D geometric awareness, 2D VLMs frequently exhibit “semantic bleeding”, where large-scale background categories (e.g., ground) erroneously submerge small-scale targets (e.g., vehicles and street elements). To address this issue, we propose a geometry-constrained pseudo-label generation and purification framework. Our approach tackles the problem through a dual-branch design: extracting open-vocabulary semantics via SAM3-based multi-view projection while simultaneously deriving sharp, class-agnostic instances using SAM2 on Gamma-transformed elevation maps. By introducing a geometric–semantic consistency module, we evaluate the internal semantic purity and external spatial homogeneity of these instances, detecting and filtering out semantic misclassifications. The purified pseudo-labels are then used to supervise a 3D sparse convolutional network via a Masked Cross-Entropy Loss. Experiments on the H3D and Turin3D datasets demonstrate that our method recovers small-scale targets that are prone to being submerged, outperforming existing zero-shot baselines by improving mIoU from 52.15% to 63.45% on H3D and from 29.52% to 58.51% on Turin3D, thereby narrowing the performance gap with fully-supervised approaches.