Cross-camera person re-identification (ReID) is a fundamental task in intelligent surveillance, yet remains challenging in realistic environments with frequent occlusions, pose variations, and complex backgrounds. The recently proposed Keypoint-Promptable Re-Identification (KPR) framework improves structural awareness through keypoint guidance, but still suffers from limited adaptive feature fusion and insufficient local context modeling under occlusion. To address these limitations, we propose Enhanced Keypoint-Promptable Re-Identification (EKPR), which integrates three complementary enhancements: (1) a Stage Attention Module (SAM) that dynamically balances multi-scale feature contributions; (2) a Keypoint-Prior CBAM (KP-CBAM) that stabilizes and refines keypoint-guided feature learning via residual attention; and (3) a deformable convolution-based pixel classification head that enlarges receptive fields and adapts to local geometric variations. Extensive experiments on the Occluded-PoseTrack21 dataset demonstrate that EKPR consistently outperforms both the baseline KPR and individual enhanced variants, achieving superior Rank-1 accuracy and mean Average Precision (mAP) across different training stages. These results confirm the effectiveness of jointly integrating structural attention, residual keypoint guidance, and flexible local modeling, providing a robust solution for person ReID under occlusion and viewpoint changes.