CMIRL: Causal Morphology-Invariant Reinforcement Learning via Mechanism-Separated World Models | Synapse