Graph Contrastive Learning (GCL) has shown promise in alleviating the oversmoothing issue in Graph Convolutional Networks (GCNs) for recommendation systems. However, existing GCL methods rely on parameter-sensitive data augmentation techniques—such as graph structure modifications or feature perturbations—that require careful hyperparameter tuning and may introduce artificial biases. To address these limitations, we propose PF-GCL++, a parameter-free framework with two key innovations. First, it introduces adjacent-layer heterogeneous node-driven view construction, which exploits the inherent structural differences between user and item nodes across GCN layers to generate contrastive views without manual augmentation. Second, it employs multi-level representation fusion, which integrates embeddings from both homogeneous and heterogeneous layers to preserve diverse structural patterns and enhance view diversity through natural neighborhood propagation. These components synergistically mitigate oversmoothing by leveraging the multi-scale nature of GCNs. Extensive experiments on three large datasets demonstrate that PF-GCL++ achieves state-of-the-art performance, particularly in sparse-data scenarios and for users with few interactions, while eliminating the need for hyperparameter tuning.