Federated semi-supervised segmentation (FSSS) has shown great potential for collaborative medical image analysis while preserving patient privacy. Despite its advantages, current FSSS methods face two main challenges. First, the scarcity of labeled data together with the non-independent and identically distributed (non-IID) nature of client data makes it difficult to generate reliable pseudo labels for unlabeled clients, which limits effective supervision. Second, aggregation imbalance across multiple clients stems from differences in label availability and data distribution among clients, causing their contributions to the global aggregation to be uneven and resulting in biased global models and degraded performance. To overcome these limitations, we propose FedCPDW, a novel framework that integrates two core strategies. First, a client-guided two-stage pseudo-label generation mechanism leverages the knowledge of labeled clients to produce high-quality pseudo-labels for unlabeled clients, ensuring more accurate supervision. Second, a dual-factor weighted aggregation strategy considers both label confidence and pseudo-label quality to balance the contributions from labeled and unlabeled clients during global model updates, thereby enhancing the robustness and generalization of the global model. We conducted extensive experiments on the multi-source HAM10K skin lesion dataset, comparing several state-of-the-art federated segmentation methods. The results demonstrate that FedCPDW consistently outperforms all baselines in both Dice and HD95 metrics, achieving higher regional overlap and more precise boundary delineation. These findings validate that FedCPDW effectively addresses the inherent challenges of FSSS and significantly improves performance in federated semi-supervised medical image segmentation.