Paper accepted to NeurIPS 2026: Latent Q-Barrier Shielding for Safe In-Context Reinforcement Learning.