Recent advances in large language models (LLMs) have created new opportunities for alleviating the long-tail problem in recommender systems by providing semantic representations that complement sparse collaborative signals. However, existing approaches suffer from two limitations. First, most fusion methods employ a globally uniform mechanism without modeling the reliability differences of items across the two information sources, making it difficult to achieve adaptive fusion for different items. Second, mutual-information-based alignment methods typically impose uniform global constraints, which, while enhancing cross-modal shared information, tend to weaken the complementarity between modalities and cause the loss of distinctive information. To address these issues, we propose a Confidence- and Divergence-aware dual-view Dynamic Fusion framework (CDDF). The proposed framework constructs user and item embeddings from both collaborative and semantic views, and adaptively adjusts the fusion weights between two embeddings based on item-specific confidence and divergence, thereby achieving complementary enhancement while maintaining modal distinctiveness. In addition, we design a knowledge distillation-based user enhancement module, which transfers knowledge from similar users to improve the robustness of preference modeling. The results show that CDDF achieves overall better performance than state-of-the-art baselines, particularly in alleviating the long-tail problem.