Reasoning Large Language Model Errors Arise from Hallucinating Critical Problem Features