Bearing fault diagnosis in industrial deployment must contend with two simultaneous distributional shifts: fault severity increases as damage progresses, and motors operate at loads unseen during training. We define this compound setting as the double domain shift and present a rigorous few-shot benchmark on the Case Western Reserve University (CWRU) and Paderborn University (PU) bearing datasets. Six architectures spanning distinct learning paradigms—a multilayer perceptron (MLP), a capsule network (CapsNet), a residual capsule network (ResCaps), a prototypical network (ProtoNet), a modified residual convolutional network (MRCN), and Deep Correlation Alignment (Deep CORAL)—are evaluated under a strict three-way split (support/validation/held-out test) that prevents the data-leakage patterns prevalent in prior CWRU protocols. Models are adapted using $K\in \{5,10,20\}$ labelled target samples and assessed on three complementary metrics: accuracy, macro F1-score, and Cohen’s $\kappa$. A lightweight 1-D convolutional neural network (CNN) with a capsule routing head (CapsNet) leads on 15 of 18 CWRU conditions and on all PU conditions at $K\ge 10$ (with MRCN leading at PU Target-A $K=5$), achieving macro F1 of $0.860$ and $\kappa =0.818$ at $K=5$ on the harder CWRU target—with 18,624 parameters (roughly one-third of the MLP baseline) and without any distribution-alignment objective. The multi-metric evaluation reveals findings invisible to accuracy alone: several baselines fall below moderate agreement ( $\kappa <0.60$) at low $K$, and MRCN’s accuracy–F1 gap of $5.3$ percentage points (pp) at $K=10$ exposes class-selective failure that accuracy conceals. On PU, a macro F1 of $0.443$ at $K=5$ on the real inner-race target quantifies the artificial-to-real fatigue transfer gap, and all models produce $\kappa <0.06$ on the outer-race target at $K=5$, establishing a realistic lower bound for future work. Wilcoxon signed-rank tests confirm the CapsNet advantage is statistically significant against the weaker baselines in nearly all conditions. Capsule output norms provide interpretable, per-class confidence-like activation scores without requiring post-hoc attribution methods.