Accurate residential energy forecasting is critical for smart grid operation and HVAC efficiency, yet it is inherently constrained by severe data scarcity and complex user heterogeneity. Standard meta-learning approaches often suffer from statistical distribution bias, as the meta-gradient typically overfits to dominant consumption patterns while neglecting minority households. To overcome this limitation, this study proposes a heterogeneity-aware model-agnostic meta-learning framework driven by a deep generative pre-categorization strategy (Cat-MAML). Abandoning heuristic statistical clustering, our approach utilizes a 1D-convolutional time-series variational autoencoder (TS-VAE) and Gaussian mixture model (GMM) soft-clustering to project raw consumption profiles into a continuous latent space, extracting profound behavioral archetypes. By enforcing a balanced task sampling distribution derived from this latent space during meta-training, the framework ensures equitable representation of both dominant and rare user groups. Experiments on a diverse, cross-climatic building energy dataset demonstrate that the proposed heterogeneity-aware framework significantly outperforms both a global baseline and vanilla MAML, achieving a 32.7% reduction in average mean squared error (MSE) and a remarkable 61.5% error reduction specifically for extreme out-of-distribution (OOD) adversarial stress-test cases. Furthermore, under simulated non-stationary online deployment scenarios, the framework exhibits superior long-term robustness and rapid adaptation efficiency, highlighting its immense practical value for scalable, data-limited residential energy management.