Multi-view clustering excels by leveraging complementary multi-perspective data, yet current methods still face the following challenges: 1) How to balance local and global representations across views; 2) How to mitigate the degradation of representation quality caused by enforcing consistency across multi-view data; 3) Many contrastive learning methods align cross-view representations but neglect clustering information, leading to false negative pairs. To address these issues, this paper proposes a framework for Progressive Feature Contrastive and Self-Weighted Neighbor-masked Fusion (PCSNFMVC). First, autoencoders extract view-specific features, where contrastive learning enforces local consistency by aligning low-level cross-view features. Then, these features are adaptively fused through a self-weighting mechanism to generate global representations, while high-level view features are projected to maximize consistency with the global representation, achieving local-global balance. Finally, a neighborhood-masked contrastive learning method is employed to distinguish high-similarity negative samples during the embedding aggregation process, effectively addressing the previously mentioned issue of false negative pairs. Experimental results on seven datasets demonstrate the superiority of this method.