ARC-AGI Without Pretraining: minuscule model (76k parameters) achieves 20% on ARC-AGI 1 with pure test-time learning, without training on the training set