Cartographic Generalization Exposes Non-Uniform Structure Loss in Downsampled Dermoscopy Benchmarks

Downsampled medical imaging benchmarks ship a reduction with no name, no contract, and no statement of what it destroyed. Cartography solved the same problem two centuries ago with named generalization operators, legends, scale bars, and source diagrams. We implement that discipline for DermaMNIST and audit the benchmark as a data readiness question. The audit is decisive before any model is trained: four of five diagnostic structure families fall below the Nyquist limit of the shipped $28^2$ grid, and the uniform resize retains structure non-uniformly across skin tone strata (spread 0.094, halved to 0.045 by equal-fidelity symbolization). We report costs in full: an 18 channel generalized sheet recovers 24\% of the rare-class recall gap and improves calibration while losing 6.8 macro AUC points; a symbol budget scale law predicted by cartographic practice does not appear; and a label-free per-image certificate's abstention value reverses between two runs of the same code. We propose four artifacts every such benchmark should ship, all machine-checkable today.