Historical optimization diagnostics for unmatched official-DINOv2, Euclid-specific SSL, and random-initialization runs on the spiral-vs.-nonspiral task. Columns show the three initialization records and rows show seeds 42–44. Curves report validation accuracy and macro-F1 over 20 epochs, vertical dotted lines mark the transition after five head-only warm-up epochs, and annotations give the best macro-F1 retained for each run. The original bottom-center frac100 panel has been replaced by the corresponding saved frac001, seed-44 history. Official DINOv2 uses replicated three-channel inputs and 384-dimensional features, whereas the Euclid-SSL and random runs use single-channel inputs and 256-dimensional features. The panels are historical training diagnostics, not a matched initialization benchmark, and they do not establish general model superiority or representation collapse.