Despite advances in graph-enhanced LLM recommenders, significant challenges remain in modeling high-order user-item dependencies. First, existing GNN-LLM hybrids propagate spurious structural correlations through indiscriminate k-hop positive aggregation, leading to ineffective representation learning. Second, LLM-based methods are often constrained by lexical surface patterns, overlooking important behavioral rationales and the dynamic nature of user preferences. To address these structural-semantic decoupling issues, we propose GNN4LMR, an enhanced recommender model that incorporates three core innovations. First, we introduce a rationale-aware profile distillation mechanism that decodes implicit interaction motives by integrating behavioral patterns with semantic contexts, producing semantically rich user and item profiles that go beyond lexical surface artifacts. This distillation reduces reliance on explicit feature mentions and better captures evolving preferences. Second, we present a graph learning mechanism enhanced with hard negative sampling, which refines sampled negatives by selecting items that are feature-close to positives but semantically irrelevant, providing stronger contrastive signals than conventional random sampling. Finally, we propose a graph-word alignment network that bridges structural graph patterns with LLM semantic spaces through adaptive neural transformations and context-sensitive fusion. Extensive experiments demonstrate that GNN4LMR consistently outperforms state-of-the-art methods in both sequential and direct recommendation tasks, showcasing its superior ability to model complex user-item relationships.