When Neural Collapse Fails to Transfer: ETF Deviation as a Diagnostic for Frozen Foundation Features

When a classifier is trained on features from a frozen foundation model in a specialized visual domain, per-class recall is not always explained by class sample size. On the PVEL-AD solar-cell defect benchmark, a linear probe on frozen DINOv2 features recalls 94% of a class with 56 labeled examples and 26% of a class with 96. We show that this inversion is explained by the geometry of the feature space. Class centroids of frozen DINOv2 and CLIP features deviate strongly from the simplex equiangular tight frame (ETF) that Neural Collapse theory identifies as the optimal class arrangement, and the deviation pattern is a property of the data domain rather than the model. Three backbones with different pretraining objectives produce near-identical per-dataset deviation ($\rho \geq 0.96$ across 11 datasets). We propose the mean pairwise centroid cosine similarity, read as an NC2 deviation statistic, as a diagnostic that separates two failure regimes. In the geometry-failure regime, centroids overlap and LoRA fine-tuning recovers accuracy by moving features toward the ETF arrangement. In the sample-scarcity regime, centroids are already separated and fine-tuning underperforms frozen linear probing. We introduce an ETF-alignment auxiliary loss derived from the diagnostic that recovers recall on confused classes while leaving overall balanced accuracy unchanged within seed variance, showing that the measured geometry is causally implicated in the failure. The diagnostic tells a practitioner, from one forward pass and roughly 10 labels per class, whether a frozen foundation model will work on their domain, and if not, whether feature adaptation or additional data is the appropriate intervention.