LLM-guided deep reinforcement learning with contrastive safety regularization for autonomous driving | Synapse