Steel fiber reinforced concrete incorporating supplementary materials offers improved toughness and durability, yet its compressive strength remains difficult to forecast because performance emerges from coupled effects of mixture composition, fiber parameters, and curing. This study compiles 446 literature-derived experimental observations of steel fiber reinforced concrete with mineral additions, records 14 input variables and compressive strength, and develops a heterogeneous stacked ensemble regressor termed HERO-R that integrates histogram based gradient boosting, extremely randomized trees, and random forest base learners with an optimized ridge meta learner, supported by feature augmentation via logarithmic transformation and z-score standardization. HERO-R is benchmarked against eight widely used regression algorithms using a five-fold cross validation protocol. The proposed ensemble achieves the highest validation accuracy with a mean R2 of 0.9405 and low prediction errors with a mean RMSE of 7.38 MPa and MAE of 4.88 MPa, while exhibiting stable generalization across folds. To complement predictive performance with interpretability, causal structure learning is used to infer a directed acyclic graph that links mixture constituents to compressive strength, including a total binder construct that causally influences the water to binder ratio and strength. Counterfactual intervention analyses, performed by fixing each variable to its minimum, mean, and maximum values and repeating the assessment across four strength strata, indicate that water content and curing duration act as dominant causal drivers across strength ranges, whereas steel fiber dosage can increase strength and fiber geometry shows limited influence within the investigated bounds. Overall, the combined predictive and causal framework provides a rigorous basis for accelerating mixture screening, reducing experimental burden, and supporting sustainability-oriented optimization of steel fiber reinforced concrete incorporating mineral additives.