Rapid estimation of building damage immediately after an explosion is essential for prioritizing inspections and supporting emergency decisions. This study develops a physics-guided, regime-switching machine-learning framework to predict building-level blast damage and derive an instantaneous resilience measure from the predicted residual performance. The dataset contains 432 building records described by 11 predictors representing blast intensity, exposure, attack regime, surrounding protection, and architectural and structural characteristics. To reduce optimistic validation caused by repeated or highly similar building profiles, the data were split at the profile-group level, producing a locked internal holdout that remained untouched during feature selection and model development. Hopkins-Cranz scaled distance was introduced as a physics-guided feature, while direct impacts were routed to a specialized CatBoost branch and all other cases to a general branch. Out-of-fold residual correction, constrained affine calibration, SHAP/ALE interpretation, and group-aware conformal prediction were incorporated into the final pipeline. On the locked holdout, the proposed model achieved MAE = 7.13, RMSE = 10.07, and R² = 0.872. Although XGBoost obtained a slightly lower overall MAE of 7.00, the difference was not statistically significant. In direct-impact cases, the regime-switching model reduced MAE from 11.21 to 8.19, indicating a clearer benefit in the safety-critical regime for which it was designed. Scaled distance, actual standoff distance, surrounding barriers, window-to-wall ratio, and TNT-equivalent mass were consistently influential. A safety-oriented conformal interval achieved 93% empirical coverage. The framework is therefore intended as a rapid screening and decision-support tool rather than a substitute for detailed structural analysis or field inspection.