Manufacturing reconfiguration in Industry 4.0/5.0 demands solutions within 20 to 50 evaluations where each experiment costs €50 to €500, while traditional methods require hundreds and evolutionary algorithms need 500 to 5,000. We present MBAE (Meta-Bayesian Active Explorer), integrating hierarchical Bayesian meta-learning with similarity-weighted transfer to leverage historical optimization campaigns. MBAE learns GP hyperparameter priors from past tasks, employs entropy-based weighting to prevent negative transfer, and uses manufacturing-aware acquisition functions. We establish sublinear regret bounds under bounded prior mismatch. Experiments across 9D problems (100 runs, 42,000 evaluations) show 4.3 to 4.6 times convergence speedup: MBAE requires only 6.16 iterations versus 26.6 to 28.3 for baselines (p < 0.0001, Cohen's d = -5.7 to -6.7). Performance gains include 2.9% to 10.3% over classical Bayesian optimization, 56.2% to 60.9% over meta-learning baselines, and 2.05× computational efficiency, confirming practical applicability for cost-sensitive industrial deployment.