Camera localization aims to estimate camera position and orientation from visual data. While recent deep learning methods formulate pose estimation as a direct regression problem, rotation is typically represented using quaternions yet estimated via standard real-valued transformations, creating a structural mismatch between learned features in Rd and the rotation manifold S3. We introduce a quaternion-aware rotation modeling approach that incorporates Hamilton-consistent interactions within the feature transformation process, replacing the conventional rotation head with a structured projection stage that explicitly couples quaternion components. Built upon CoHAtNet, our method achieves consistent improvements in median rotation error across all seven scenes of the 7-Scenes dataset, reducing the average rotation error from 0.74° to 0.67°, while maintaining competitive translation accuracy.