The paper presents MSGCA (Multimodal Stable Fusion with Gated Cross-Attention), a novel framework for fine-grained stock movement prediction that integrates three modalities: numerical indicator sequences (e.g., prices), textual dynamic documents (news and tweets), and a relational graph (industry sectors). The framework consists of three phases: (1) a trimodal encoding module using MLP, a pre-trained LLM, and GAT to project each modality into a unified latent space; (2) a stable fusion module employing two gated cross-attention networks, where the primary modality (indicators) guides the sequential integration of document and graph features, effectively mitigating data sparsity and semantic conflicts; and (3) a prediction module that reduces temporal and feature dimensions via MLPs to output probabilities for three movement trends: up, flat, and down. Empirical evaluations on four real-world datasets demonstrate that MSGCA significantly outperforms existing unimodal, bimodal, and multimodal baselines, achieving accuracy improvements of 8.1% to 31.6%. The gated cross-attention mechanism is key to stable fusion, as it leverages the most reliable modality to filter noise from others, addressing the inconsistency challenges common in multimodal financial data. This work fills a gap in fine-grained multimodal stock prediction and offers a robust solution for practical investment strategies.