Expressive Score-Based Priors for Distribution Matching with Geometry-Preserving Regularization