Chinese spelling correction faces unique challenges due to the multi-dimensional nature of Chinese characters encompassing phonetic, visual, and semantic properties. Existing methods treat phonetic and visual features as auxiliary information, limiting their effectiveness. We propose CSC-FGMF (Chinese Spelling Correction via Fine-Grained Multi-Feature Fusion), a novel framework that introduces fine-grained multi-feature fusion through three key innovations: (1) Error-aware adaptive fusion employs cross-modal attention to dynamically weight features based on predicted error types, enabling specialized correction strategies; (2) Graph-enhanced IDS modeling uses graph convolutional networks to process character decomposition trees, capturing structural relationships and visual similarities; (3) Context-sensitive confusion learning generates position-specific candidates by integrating linguistic priors with contextual representations. Our framework systematically integrates four complementary representations: BERT semantic embeddings, hierarchical phonetic decompositions, graph-based structural features, and dynamic confusion sets. Experiments on SIGHAN benchmarks show significant improvements: 93.1% F1-score for character-level correction on SIGHAN 2015 (5.0% absolute improvement) and 90.6% for sentence-level detection. Ablation studies confirm that multi-feature fusion contributes 17.2% improvement, with each component providing substantial gains across diverse error patterns.