The management and analysis of large-scale, heterogeneous time series datasets pose significant challenges to traditional monolithic forecasting models due to the diverse underlying dynamics. While Mixture-of-Experts (MoE) enables specialization, existing approaches suffer from three critical gaps: task-agnostic clustering optimizing generic similarity; deterministic routing ignoring prediction uncertainty; black-box pattern discovery lacking causal interpretability. We propose ClusterPatchTST with three synergistic innovations: gradient-guided cluster evolution, automatically discovering optimal pattern granularity via forecasting gradient analysis; uncertainty-aware expert routing, Bayesian-principled down-weighting of unreliable predictions; and causal-aware pattern discovery, learning directed acyclic graphs (DAGs) characterizing cause-effect relationships within each pattern, enabling mechanistic interpretation and out-of-distribution generalization. Through end-to-end joint optimization with KL-divergence alignment and contrastive learning, our method achieves robust task-aware specialization. On standard benchmarks, ClusterPatchTST sets a new state-of-the-art, reducing average MSE substantially over its PatchTST backbone and MoE baselines. Gradient-guided evolution discovers several interpretable clusters; uncertainty routing improves 6.1% on noisy data; causal discovery achieves better out-of-distribution generalization, with learned causal graphs revealing actionable mechanistic insights.