Reverse distillation has achieved impressive performance in the challenging domain of unsupervised anomaly detection. However, one significant limitation of this approach is the potential for insufficiently compact feature reconstructions, which can lead to an increased rate of false positives, particularly when identifying subtle defects in confectionery products. To address this critical issue, we propose a novel method called Recompacting Reverse Distillation via Hypersphere (RRDH). At the heart of our approach is the Hypersphere Compactness (HC) module, which integrates a hyperspherical geometry into the adapter network. This integration further compresses the feature embeddings that the student decoder must reconstruct, ensuring their sufficient compactness and thereby reducing the likelihood of false positive detections. Extensive experiments conducted on the Eyecandies-2D dataset, a specialized benchmark for confectionery anomaly detection, have shown that RRDH achieves state-of-the-art performance. Our results not only validate the effectiveness of the proposed method but also underscore its theoretical soundness.