Leakage-Aware Retrofit Monitoring for Collaborative Robot Anomaly Detection

External inertial sensing enables controller-independent monitoring of collaborative robots. Robust evaluation should distinguish classifier discrimination from temporal post-processing and explicitly control temporal leakage, sensor–label misalignment, and installation-sensitive sensing shortcuts. This study investigates the UR3e/Nicla Sense ME record using complete-event chronological separation, transition masking, training-only transformations, same-split sensor-channel ablation, raw per-window Random-Forest (RF) scores, and validation-selected H-of-K persistence. Two velocity-anomaly episodes are excluded from all model, threshold, channel, and persistence selection. The same-split ablation selects a gyroscope-only 31-feature RF representation. Before persistence, the held-out RF yields TN = 153, FP = 231, FN = 0, and TP = 384 (MCC = 0.4988), showing high anomaly recall but poor nominal specificity. Validation selects a 5-of-5 persistence rule, yielding TN = 382, FP = 2, FN = 8, and TP = 376, with accuracy = 0.9870, precision = 0.9947, recall = 0.9792, F1-score = 0.9869, and MCC = 0.9741. Both held-out anomaly episodes are detected with an approximately 40 s confirmation delay after the first eligible positive window. The 21-of-21 rule is retained as a consistency reference and reproduces TN = 384, FP = 0, FN = 40, and TP = 344, with approximately 200 s confirmation delay. Guard-band, persistence, sensor-channel, rolling-origin, protocol-component ablation, and same-split comparator analyses show that performance depends on the evaluation protocol and held-out event. The evidence supports bounded later-event recognition within the same robot and acquisition domain; future studies should establish the long-term false-alarm rate and test cross-platform generalization.