Quality-sensitive, energy-intensive manufacturing systems face quality-dependent capacity, time-of-use energy conditions, and customer-specific acceptance rules. This paper develops a learning-driven optimization framework that converts a surrogate-predicted quality–capacity Pareto frontier into a finite library of candidate operating recipes and embeds it in a recipe-coupled mixed-integer linear programming (MILP) model. Each recipe carries a predicted quality level, an achievable-capacity estimate, and operating set-points; the MILP selects recipes over a weekly hourly horizon while coordinating production, inventory, shipments, shortages, time-of-use tariffs, electricity availability, and customer quality compatibility. The framework is demonstrated on a rotary-kiln lime case study in which loss on ignition is the quality metric: the surrogate models are developed from historical kiln PLC records, while the planning experiments use a reproducible generated reference instance parameterized by a 20-recipe subset of the archived Pareto candidates. For the reported instance (11,257 constraints, 8,444 variables, 3,024 binary), the MILP is solved with HiGHS in about 48 s to within the solver’s default 10⁻⁴ relative optimality gap. Against fixed-capacity and no-coupling baselines, the value of recipe capacity coupling is negligible under slack capacity but reaches up to 24% of weekly profit near full kiln utilization. Demand service is analyzed under a shortage-penalty sweep and a mandatory-order stress case, the surrogates are re-evaluated under chronological lag- and rolling-window tests, and all Pareto candidates are screened against the historical operating envelope; the corresponding evidence boundaries are reported explicitly. A bounded supervisory layer may update only approved scenario data and trigger re-solving of the unchanged model.