v3: deep clean via per-source cross-val + manual curation

add per-label verification composites and cleaning scripts

cleanup: remove temp files, fix find_label_issues.py guard

fix: build_dataset.py writes HF Image feature type

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