From Noise to Laws: Regularized Time-Series Forecasting via Denoised Dynamic Graphs