nx_de_referee_gate.nx

nx_de_referee_gate.nx -- THE GATE FOR THE PER-REGION CIEDE2000 REFEREE, 2026-09-03. SUBJECT: dr_region_de / dr_run in-process (nx_de_referee_lib has no main). THE THREE TEETH THAT CARRY THE CLAIM ARE T4, T5 AND T9, and every one of them is a law the estate has already been bitten by: T4 ZERO REGIONS MUST NOT PASS. A tooth that passes on the empty set is not a tooth, and a referee that returns PASS when it measured nothing is the most dangerous possible green. T5 ALL-UNOBSERVED MUST NOT PASS either -- declaring regions and sampling none of them is the same vacuum wearing a denominator. T9 AN UNOBSERVED REGION RESCUES NOTHING. If one measured region is out of tolerance, adding empty regions beside it must not turn the verdict green. That is the abstain-never-acquit law in the one direction where getting it wrong is silent. Teeth, in order: T1 identical colours give a difference of exactly zero. T2 a real difference is non-zero (so T1 cannot pass on a stub that always returns zero). T3 the WORST region is correctly identified, not the first or the last. T4 zero declared regions -> ABSTAIN, never PASS. T5 all regions unobserved -> ABSTAIN, never PASS. T6 the partition RECONCILES: measured + unobserved == declared, always. T7 a render inside tolerance PASSES. T8 a render outside tolerance FAILS -- with T7 this proves the verdict is not a constant. T9 an unobserved region does not rescue a failing measured one. T10 NEG-CONTROL: a non-positive tolerance REFUSES rather than judging against nothing. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0