Few-shot meta-learning supports rapid adaptation with limited labeled data, but remains sensitive to hyperparameters and support set quality. We propose Uncertainty-Decomposed Reliability-Gated Meta-Learning (UDRML). It selects support sets with low predictive entropy, combines diverse adapted models to estimate aleatoric and epistemic uncertainty through mutual information, and uses a reliability score to accept, flag, or reject predictions. We evaluate UDRML on cross-domain few-shot LiDAR fault classification using a synthetically augmented nuScenes dataset. At 10-shot, UDRML improves accuracy over MAML by 11.8 points. Gating provides a further 13.3-point gain on accepted predictions, while measured epistemic uncertainty is 78% lower than that of the strongest ensemble baseline, ECMP. The framework supports selective automation by accepting reliable predictions and directing uncertain cases to verification or a fallback.