Beyond Sparse Rewards: Self-Supervised Structural Reinforcement Learning via Combinatorial State Restoration | Synapse