Medical image segmentation critically relies on pixel-level annotated data, yet its prohibitively expensive acquisition necessitates efficient semi-supervised learning strategies to alleviate data dependency. Contemporary methods remain constrained by either single-level prototype modeling or stochastic augmentation, inadequately capturing complex anatomical details and yielding noise-sensitive pseudo-labels. To address this, we propose a novel semi-supervised framework integrating a Hierarchical Prototype System (HPS) with Adaptive Multi-position Copy-Paste (AMPCP) augmentation. HPS jointly models global semantics and substructural features through dual-level (L1/L2) prototypes enhanced by EMA-driven uncertainty guidance, improving adaptability to anatomical variations. AMPCP employs complexity-aware scheduling with intelligent position generation and dynamic weight adjustment for contextually adaptive hybrid augmentation. The synergistic integration of HPS and AMPCP generates highly reliable pseudo-labels, significantly boosting training stability and generalization. Extensive experiments demonstrate state-of-the-art performance across multiple medical segmentation benchmarks, validating exceptional efficacy in low-annotation scenarios.