Multi-view clustering focuses on extracting complementary information from multiple data modalities to uncover latent semantic structures. While existing methods typically rely on structural similarity for view fusion, they often overlook weak supervision signals such as pairwise constraints, resulting in insufficient exploitation of semantic information. Moreover, they are prone to semantic collapse due to excessive inter-view alignment, where views become overly similar and lose discriminability. In pursuit of this objective, we introduce Constraint-Aware multi-view clustering via Graph Contrastive Learning (CAGCL), which is designed to incorporate weak pairwise supervision and mitigate semantic collapse through structure-aware contrastive learning. We encode pairwise constraints into view-specific graph structures through graph contraction and Laplacian adjustment. To effectively integrate pairwise constraints and address semantic collapse, CAGCL incorporates two key components: a structure-aware contrastive objective and a dynamic diversity regularization. A structure-aware contrastive objective preserves constraint-induced semantics by aligning similar samples in the embedding space. Meanwhile, a dynamic diversity regularization promotes view-specific discrimination and mitigates semantic collapse. Together, these components form a unified framework that balances consistency and diversity under weak supervision. Extensive experiments on eight benchmark datasets show that CAGCL achieves superior performance on par with state-of-the-art methods.