Avoiding Over-Personalization with Rule-Guided Knowledge Graph Adaptation for LLM Recommendations