“Dexterous manipulation policies need dense, richly annotated human-to-robot retargeting examples across everyday domains — not just isolated single-task demonstrations.”