Drop 37 ambiguous/badly-cropped samples where model disagrees with label

Clean labels using Qwen3-VL-4B VLM verification

v2: add new sources, extraction pipeline, better train/val split

Update README with full dataset documentation

Initial dataset: racing gear digits + MNIST supplement