Social recommendation systems leverage user social connections to enhance personalized recommendations. However, existing methods face challenges in fully capturing the distinctions between semantically similar users and items, which can result in redundant embeddings and limit recommendation diversity. This also affects the quality of recommendations for long-tail items. Moreover, conventional loss functions remain static, which can lead to suboptimal model convergence. To address these limitations, we propose Reinforcement-Optimized Contrastive Graph Representation Learning (RCGRL), a framework integrating contrastive learning and reinforcement learning. Specifically, its contrastive learning module features a Positional Weight InfoNCE loss which utilizes hybrid similarity rankings to explicitly maximize distinctions among semantically related entities. This significantly enhances embedding discriminability and robustness. Concurrently, an actor-critic reinforcement learning mechanism adaptively adjusts loss function weights. Crucially, this dynamic optimization is activated based on a stability assessment derived from real-time performance fluctuations, allowing the model to continually optimize its learning process guided by recommendation performance metrics. Extensive experiments on two real-world datasets demonstrate the effectiveness of RCGRL.