Consistency and version regression checks

Reordering and paraphrase are the same check with a different transformation: ask twice, compare top answers by label, and list the flips with their probabilities. The regression half compares two pinned versions item by item: paired_table() counts the four cells, sign_test() is the exact two-sided McNemar test on the discordant cells, and check() blocks the switch on any class recall drop over 5 points or an ECE rise over 0.02, whatever the overall accuracy. The offline run uses two deterministic simulated "versions" with a built-in lean towards the first-listed option, so every check has something to find.