Multi-Modal Semantic Expansion with Constrained LLM Reranking for Conversational Music Recommendation