Reliable milling monitoring requires calibrated intervention probabilities and evaluation that respects experimental clustering. This methods-and-evaluation study examines Hierarchical Bayesian Adaptive Fusion (HBAF), combining calibrated TabM and Gaussian Naive Bayes experts. MSM provides 927 source-machine cutting paths and 1585 from a historically inspected second machine. Nested development assessment uses 563 paths from 11 experiments; frozen evaluation additionally uses 172 internal-test paths and the second-machine cohort. The development of HBAF achieves MCC 0.6383, PR-AUC 0.8552, and Brier score 0.0773, but equal fusion and a nonhierarchical control remain stronger on several measures. Added component, prefix, scale-sensitivity, and NUTS analyses do not establish general hierarchical-fusion superiority or uniformly valid Laplace uncertainty. Frozen HBAF performance decreases to MCC 0.4587 and PR-AUC 0.5820 internally, and MCC 0.1040 and PR-AUC 0.2809 across machines. The experiment-aware inference also leaves asymmetric-cost superiority unestablished. The contribution is a leakage-safe evaluation framework and a reproducible case study identifying limits of Bayesian adaptive fusion. Historically inspected transfer data, conditional posterior uncertainty, and hypothetical error costs do not support claims of independent cross-machine confirmation or industrial deployment validation.