Depth-Blind by Design: An Interpretability Autopsy of a Frozen Kepler Transit Classifier on TESS

Convolutional classifiers trained on Kepler light curves are routinely reused on TESS, where they rank candidate planets at a scale no astronomer can inspect by hand. When such a pipeline returns nothing, the aggregate score cannot say whether the sky is empty, the search failed, or the network is blind. We open one up. Three interpretability probes, applied to a frozen classifier deployed on three TESS populations, each yield knowledge the domain did not have. An ablation over ten seeds localises essentially all discriminative signal in the local view: an 8.4M-parameter global branch, 86% of the network, buys at most 0.005 PR-AUC (95% CI $[-0.004, +0.005]$, Welch $p = 0.83$). A per-stage decomposition of 417 missed confirmed planets shows the network, not the rule-based vetting, is the binding constraint (26.0% against 21.6%; McNemar $p = 0.007$), and that this survives conditioning on a correct search ephemeris. Finally, a behavioural probe explains a near-null on 474 faint stars that looks like a sensitivity limit and is not: 41.6% of search periods fall within 10% of the 13.7-day TESS orbital period, a 7.6-fold excess over chance, and the classifier is indifferent to them, scoring 0.030 against 0.032 elsewhere. The cause is not learned but architectural. Depth normalisation divides every view by its own minimum, so the deepest bin of each view is exactly $-1$ and absolute transit depth is destroyed before the network sees it, leaving the one feature that would expose these artefacts provably absent from the representation.