Accurate volumetric quantification of brain tumor subregions in magnetic resonance imaging (MRI) is sensitive to local prediction bias, because boundary over-expansion, missed enhancing foci, and class confusion at subregion transitions can propagate to case-level volume errors. To address this issue, this study proposes CPLR-Net, a coarse prior-guided logit-space residual refinement framework with functional symmetry between global layout preservation and local error compensation. A Coarse Prior Segmentor first generates coarse logits that encode the global semantic layout of the tumor. These logits are then concatenated with multi-modal MRI and fed into an efficient multi-scale attention (EMA)-enhanced Residual Refiner, where three-dimensional cross-spatial feature recalibration and gated logit residual updating are used to correct ambiguous boundaries and subregion transitions while maintaining consistency with the coarse semantic prior. Experiments on the Brain Tumor Segmentation (BraTS) 2020 and 2021 datasets show that CPLR-Net achieves favorable overall performance in the Dice similarity coefficient, 95th percentile Hausdorff distance, and absolute volume difference compared with representative three-dimensional segmentation methods, supporting accurate segmentation and volumetric quantification. The results suggest that coarse-prior-constrained residual refinement in logit space provides an effective strategy for reliable key-subregion segmentation and quantitative assessment in brain tumor MRI follow-up scenarios.