v2.1: the benchmark becomes its difficulty-selected form - 1,696 items solved by at most 6 of a frozen 7-model pool; manifest and per-item difficulty shipped

Difficulty-curated release: 1,696 items (pass<=6/7 against the frozen 7-model pool); manifest included

Replace the blind-probe build with the closed-book knowledge build: 2,665 items

Formosa Bench v2: 894 Taiwan items from Taiwan.md, blind-probe filtered

initial commit