In real-world clinical settings, deploying medical AI applications requires lightweight models that can operate under limited computational resources. For skin lesion segmentation, a crucial step in the early detection of skin cancer, the key challenge is to develop models that are not only efficient but also perform reliably with minimal annotated data. To address this, we propose a lightweight and efficient semi-supervised segmentation framework that combines multi-task consistency learning with the representational power of foundation models. Our method is built on three key components: (1) a dual-network co-training framework combining a lightweight MobileNet with a strong ViT-based teacher to balance efficiency and representation power, (2) a fused mask prompt inspired by multi-task consistency, which combines coarse segmentation masks with boundary-aware Signed Distance Function (SDF) maps to guide SAM, and (3) a SAM-guided knowledge distillation strategy, where refined outputs from SAM are used as high-quality pseudo-labels to train the Main Network on unlabeled data. Extensive experiments demonstrate that our approach achieves competitive segmentation performance with significantly reduced annotation effort, offering a practical solution for semi-supervised medical image segmentation in real-world applications.