A Multifidelity Sim-to-Real Pipeline for Verifiable and Compositional Reinforcement Learning