Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents