Robust or Suggestible? Exploring Non-Clinical Induction in LLM Drug-Safety Decisions