Motor imagery (MI) electroencephalography (EEG) decoding remains challenging due to severe cross-subject variability and signal non-stationarity, which significantly degrade model generalization to unseen subjects. Existing prototype-based and domain adaptation approaches typically model EEG representations at a single semantic level, limiting their ability to simultaneously capture spatial channel characteristics and high-level discriminative structures. To address this limitation, we propose a dual-stage prototype representation framework for cross-subject MI-EEG decoding. Specifically, electrode channel prototypes and feature prototypes are jointly constructed to enable hierarchical representation learning at both spatial and semantic levels. On this basis, a prototype-guided pairwise similarity learning mechanism is developed, where label supervision in the source domain and a pseudo-supervised structure in the target domain explicitly regularize inter-sample semantic relationships, improving intra-class compactness and inter-class separability. Furthermore, a StyleMix-based feature perturbation strategy and a Wasserstein distance–driven domain alignment module are incorporated to mitigate cross-subject distribution discrepancies and enhance feature robustness. Extensive cross-subject experiments on public brain–computer interface (BCI) competition datasets demonstrate the effectiveness of the proposed method. On the BCI Competition IV dataset 2a, our approach achieves an average classification accuracy improvement of 2.99% over state-of-the-art methods, and on dataset 2b, the improvement reaches 2.83%. These results consistently validate the superiority of our method across multiple evaluation metrics.