Assessing the Limitations of Large Language Models in Clinical Fact Decomposition