arXiv:2606.20197v4 Announce Type: replace Abstract: Hybrid optimization-learning frameworks, such as Actor-Critic Model Predictive Control (MPC), rely on dynamically adjusting cost parameters to manage complex control landscapes. In this paper, we introduce a certifiably robust Transformer-Actor-Critic MPC architecture that leverages history-aware sequence models to shape local optimization while maintaining strict physical safety guarantees. First, we mathematically constrain the Transformer sequence models to satisfy global incremental Input-to-State Stability ($\delta$ISS). Second, we integrate these neural Lipschitz bounds with Riemannian contraction theory, deriving a novel coupled small-gain condition that enforces uniform closed-loop contraction stability across the interconnected system. Evaluated on quadrotor control and decentralized multi-agent navigation tasks subjected to severe observation noise and process disturbances, our closed-loop stable formulation successfully translates theoretical bounds into physical safety.