Accurate and efficient analysis of electrocardiogram (ECG) signals is essential for early detection of cardiac arrhythmias. However, many deep learning models suffer from limited generalization to unseen patients, reduced interpretability, and high computational demands. To address these challenges, we propose HAT-ECG, a hybrid Autoencoder-Transformer architecture that integrates unsupervised feature learning with attention-based temporal modeling. The convolutional autoencoder extracts compact, noise-robust latent representations of ECG beats, while the Transformer's Multi-Head Attention mechanism adaptively focuses on diagnostically relevant waveform segments. The model was evaluated on three public datasets: MIT-BIH, INCART, and the independent PTB Diagnostic ECG Dataset (290 subjects, entirely outside the original datasets). Under standard beat-wise splitting for benchmarking, HAT-ECG achieved state-of-the-art accuracies of 99.91% (MIT-BIH 5-class), 99.69% (MIT-BIH AAMI), 99.15% (INCART 3-class AAMI), and 98.45% (PTB 2-class). Critically, under strict patient-wise splitting on MIT-BIH (completely disjoint training and test patients), the model maintained a realistic accuracy of 90.81% (F1-score 92.61%), with strong generalization to new patients. With only 0.021 GFLOPs, HAT-ECG offers an excellent balance of high performance, interpretability, and efficiency, making it highly suitable for real-time wearable and edge-device cardiac monitoring. This work advances deep learning for intelligent and deployable arrhythmia classification by combining accuracy, cross-patient generalization, and computational minimalism.