GALATEA II: Benchmarking LLM Safety in Clinical Simulation. Behavioural Safety and Ethical Robustness of Large Language Models in a Multi-Agent ICU Decision Support Architecture | Synapse