Class Activation Mapping (CAM) variants are widely used to explain deep learning segmentation models, yet it remains unclear whether visually similar variants respond consistently to controlled anatomical perturbations—a property we term counterfactual consistency, which is a prerequisite for trustworthy clinical deployment. We present Synthetic Counterfactual Border Audit++ (SCBA++), a framework that audits attribution consistency by combining thin-plate-spline (TPS) border warping with Poisson blending to generate anatomically plausible lungboundary perturbations, and then measures how four decoderfocused CAM variants respond via three metrics: attribution mass change in the region of interest (∆AM ROI), center-ofattribution shift (CoA Shift), and directional consistency (DC). Experiments on the JSRT (247 images, 38 test cases) and Montgomery (21 cases) chest X-ray datasets with a UNet segmentation model show that Multilayer CAM achieves significantly smaller attribution centroid shifts than Grad-CAM++ (JSRT: 4.55 vs. 6.03 pixels, Bonferroni-corrected paired t test p = 6.0 × 10−5 , paired Cohen’s dz = −0.73; replicated on Montgomery with dz = −0.88). Directional decomposition reveals positive mean responses across all methods, confirming net tracking of boundary edits, while CoA Shift exposes a sensitivity–stability tradeoff: methods with higher attribution-mass responsiveness exhibit greater spatial instability. These results demonstrate that CAM variants are not interchangeable under clinically meaningful perturbations and that counterfactual consistency provides a complementary auditing dimension beyond static localization metrics.