nx_reader_ruler_adversary_gate.nx

nx_reader_ruler_adversary_gate.nx -- ADVERSARY + CRITIC pass on the READER F1 RULER, independently confirming (or discrediting) the 2026-07-13 finding: "the 2-shot surface-form win (+104 on the single-gold standalone bench) was a METRIC ARTIFACT -- on the multi-alias on-protocol metric it did NOT transfer (SQuAD 427 vs 510), so it was reverted." Per the sovereign critic-kernel discipline (see nx_card_adversary_gate): a RULER is trustworthy (LAW) only if it DISCRIMINATES genuine-correct from an adversary built to game it. Here two rulers are compared: single_gold_f1(pred, gold) = qs_f1 against ONE gold string (the standalone bench metric) multi_alias_f1(pred, aliases[]) = max_i qs_f1(pred, alias_i) (the on-protocol qab metric == official SQuAD) The ADVERSARY = a surface-form change ("7" -> "seven") that games single-gold WITHOUT improving true correctness. A trustworthy ruler must give the adversary ZERO gain. We prove: single-gold IS gamed (delta>0), multi-alias is NOT (delta==0). That mechanically CONFIRMS the revert. license_tier: ORIGINAL expect_exit: 0