This paper conducts bottleneck merging control experiments and empirical analyses of anomalous events under varying traffic demand densities and heterogeneous driving behaviors, and establishes a Safety-enhanced and Attention-guided Multi-Agent Proximal Policy Optimization (SA-MAPPO) collaborative control model. Based on this model, a self-attention mechanism and a dual-layer safety constraint comprising action masking and Responsibility-Sensitive Safety (RSS) are adopted in place of conventional multi-agent reinforcement learning algorithms, thereby achieving coordinated optimization of driving safety, ride comfort, and system-level traffic efficiency without compromising vehicle throughput. Validated under high-density demand and mixed-fleet conditions, the SA-MAPPO-controlled bottleneck exhibits a peak collision rate suppressed within 0.80%, a merging success rate maintained above 99.2%, and an average vehicle jerk reduced to within 0.85 m/s3, demonstrating the high environmental adaptability and robustness of the proposed collaborative control algorithm. Furthermore, a causal saliency scoring model established based on this architecture successfully identifies the source perturbing vehicle, effectively overcoming the "victim-punishment" shortcoming, and provides a solid theoretical foundation for proactive traffic management and anomaly de-escalation.