nx_lang_h2h_gate.nx

nx_lang_h2h_gate.nx -- THE REFEREE for the /compare/lang head-to-head receipt (2026-09-01). Subject: knowledge/compare/lang.bench, the receipt nx_lang_h2h writes, read through the ONE reader every consumer shares (nx_bench_receipt_lib). The gate proves the receipt is a measurement and not a claim: - a reference arm and a NishiLang arm exist and are VALID, every VALID arm printed the reference checksum, every VALID arm carries its run count, min <= median, a positive median, a 64-hex source sha256 and a 64-hex binary sha256, a ratio derived from the reference median, a repro command and a 4-field host line; - the writer's @verdict and the reader's re-derived verdict AGREE (rows edited after writing are caught); - the classifier BITES: a planted checksum mismatch reads VOID, a missing reference reads UNMEASURABLE, an empty receipt reads EMPTY, a forged @verdict over void rows is a writer/reader disagreement. It says NOTHING about whether the numbers are good: that is the page's honest verdict to state and the operator's to accept. A receipt that is absent SKIPs (I could not look), never RED. exit: 0 GREEN . 1 RED . 3 SKIP (gv_verdict) license_tier: ORIGINAL No hw writes (Rule 26).