When More Metadata Misleads: Probing Field-Level Grounding in LLM-Generated Museum Descriptions