Weakly-supervised candidate generation is common practice in two-stage approaches for pulmonary nodule detection, as it reduces annotation costs. We describe a systematic failure mode of pseudo-label training in this setting, which we term boundary bias: 3D CNNs tend to assign high detection confidence to the periphery of CT volumes, rather than true nodule locations. This leads to substantially reduced recall, while standard evaluation metrics, including the loss on pseudo-labels and the patch-level ROC-AUC, are only mildly affected, thus masking the issue. We observe this phenomenon on the LIDC-IDRI dataset (888 volumes, 1186 nodules) and hypothesize that it is caused by the network learning an easy shortcut of detecting the air-tissue interface, encoded in the HU gradient. We then propose a four- step Spatial Audit Protocol for sliding-window pipelines, which can be used to evaluate any candidate generation method. We find that GT-supervised retraining reduces boundary bias (CPM: 0.053→0.194, ×3.7 improvement). Similarly, a lung-masking ablation reduces, but does not eliminate, the bias. Finally, we find that the choice of window stride implicitly controls the maximal sensitivity via an upper bound on the number of candidates, independent of the classifier performance.