Sleep quality is central to human health, yet reliable and scalable sleep assessment remains an unmet challenge in both clinical and home-care settings. Manual scoring is labor-intensive and impractical for long-term monitoring, whereas existing automatic approaches often lack computational efficiency, deployability, and interpretability. Here we present SomnoNet, a domain-informed neural architecture that unifies accurate, lightweight, and interpretable sleep staging. SomnoNet is an end-to-end framework that learns directly from raw single-channel EEG, eliminating hand-crafted preprocessing and achieving state-of-the-art performance on two large-scale benchmarks (80.9\% accuracy on SHHS; 88.0\% on Physio2018). We further develop SomnoNet-Nano, a highly compact variant that achieves an extreme parameter reduction-approximately 6\% of the smallest prior model-while still preserving more than 99\% of state-of-the-art accuracy, thereby enabling deployment on portable and wearable devices. To promote clinical trust, we conduct interpretability analyses that quantify the contribution of EEG features across epochs, exposing physiologically meaningful patterns that reveal the network's decision process. By jointly addressing accuracy, efficiency, and transparency, SomnoNet provides a practical pathway toward reliable and scalable AI-driven sleep assessment. The implementation is publicly available at https://github.com/komdec/SomnoNet.git.