Add manifest (metadata.json)

metadata: 2026-07-07 revised taxonomy (task_type format vocabulary multiple_choice/yes_no/short_answer/free_form; score_type incl. none; re-verified protocol evidence; repaired prompt provenance)

metadata: add score_type/score_params/score_protocol (official grading protocol taxonomy); task_type is now format-only (judge_scored migrated to vqa + llm_judge). See dataset-upload skill references/metadata-json.md.

Tighten prompt_template_source.notes: explicit 5-tier classification per 2026-06-01 priority rules

Refine prompt_template_source: lmms-eval/VLMEvalKit canonical citations

Populate prompt_template_source for all subsets per 2026-05-30 audit

Add v2 manifest (metadata.json)