Adaptive GR(1) Specification Repair for Liveness-Preserving Shielding in Reinforcement Learning