<p>Effective urban traffic control is essential not only for improving travel reliability and reducing vehicle emissions, but also for enhancing public safety. Reinforcement learning (RL) has shown strong potential for optimizing adaptive traffic signals. However, most existing RL-based systems lack the flexibility to prioritize emergency vehicles such as ambulances and fire trucks without significantly compromising regular traffic flow. To address this, we propose a dual-policy multi-agent reinforcement learning framework for context-sensitive traffic signal control that balances emergency responsiveness with overall efficiency. Our method employs a real-time contextual flag to switch between two specialized control policies: one for regular traffic and another for emergency scenarios. During emergencies, the system dynamically reconfigures signal phases to facilitate the smooth passage for high-priority vehicles. In the absence of emergencies, decentralized intersection-level agents use deep RL to minimize congestion and optimize traffic flow. Experiments conducted on two road networks demonstrate that the proposed framework effectively reduces both travel and waiting times for emergency and regular vehicles.</p>