LLMs have shown promise in legal NLP, but generating legal reasoning that is factually grounded, faithful to precedent, and consistent with judicial analysis remains difficult. Retrieval-augmented generation is often used to improve grounding, yet standard legal RAG faces two limitations: retrieved materials are hard to synthesize, and retrieval is driven by surface similarity rather than the latent legal issues and factual considerations that govern the analysis. We address these limitations with SCALE (Structured Case-law-guided Anticipatory Legal reasoning), a framework for precedent-grounded legal reasoning. SCALE converts article-specific Case Law Guides into an Issue–Factor Hierarchy, identifies the issues and factors implicated by a case, retrieves the linked guide passages, and generates the final legal analysis. Across Llama-3.1 and Qwen3 models, expert evaluation shows that SCALE consistently outperforms prompt-only and retrieval-based baselines, with the largest gains in legal issue identification and precedent use. These gains are especially strong for smaller models, suggesting that explicit legal structure can compensate for weaker implicit reasoning capacity. We also compare expert assessment with an LLM-as-a-Judge framework and find only weak-to-moderate agreement, underscoring the limits of automated evaluation for high-stakes legal reasoning.