Accurate forecasting of typhoon-induced waves is critical for coastal hazard mitigation and ensuring the safety of offshore operations. However, existing state-of-the-art deep learning approaches typically struggle to capture extreme wave heights under typhoon conditions, limiting their reliability in extreme disaster risk assessment. To address this bottleneck, this study constructs a physics-guided deep learning model for significant wave height in the Western Pacific, using the Cuboid-Transformer framework and optimizing it via an extreme-error loss optimization strategy. Regarding physical constraints, we explicitly decompose the wave spectrum into wind sea and swell components, guiding the model to learn the distinct evolutionary dynamics of locally generated wind waves versus remotely propagated swells. To resolve upper-tail underestimation, we integrate a high-error penalty term directly into the loss function. This mechanism forces the model to assign an order-of-magnitude higher weight to errors in peak wave heights during training, thereby purposefully enhancing forecasting accuracy at the upper tail of the distribution. Taking the recent Typhoon Bavi during July 2026 in the Northwest Pacific as a case study, we employed a 48-hour rolling forecast scheme driven by hourly GFS data and validated the results against ERA5 reanalysis. The model yielded an RMSE of 0.278 m and a CC of 0.936, enabling 48-hour forecasts with an inference latency on the order of minutes.