nx_de_referee_lib.nx

nx_de_referee_lib.nx -- THE PER-REGION CIEDE2000 REFEREE: rung ST1 of /compare/skintwin (2026-09-03) WHY THIS IS THE HIGHEST-VALUE RUNG ON THAT BOARD. nx_colorsci_lib has carried cs_de2000 -- the colour difference the whole field states its errors in -- since 2026-08-24, and MEASURED with corpus_complete=1 it has ZERO adopters outside the garment-print lane. Until something in the render lane calls it, every quality claim about a rendered human is an assertion about MECHANISM, not about appearance. The chromophore stack can be perfectly derived and still look wrong, and nothing in the estate could say so in a number. This file is the thing that can say so. IT COMPOSES, IT DOES NOT RE-RULE. The colour difference itself is cs_de2000, untouched: this file adds only the AGGREGATION -- per region, worst region, mean, and a verdict against a declared tolerance. There is exactly one CIEDE2000 in this estate and it is not here. THE THREE LAWS THIS FILE EXISTS TO ENFORCE, each of which the estate has been bitten by before: 1. ABSTAIN, NEVER ACQUIT. A region with no samples is UNOBSERVED. It is not a pass. A referee that silently counts an unmeasured region as agreement is worse than no referee, because its green is indistinguishable from a real one. 2. A TOOTH THAT PASSES ON THE EMPTY SET IS NOT A TOOTH. If nothing was measured, the verdict is ABSTAIN. Zero regions must never return PASS, however tight the tolerance. 3. THE DENOMINATOR TRAVELS WITH THE VERDICT. measured + unobserved always equals the declared region count, and both halves are written to the output so a caller can reconcile them rather than trust a bare number. The TOLERANCE is an ARGUMENT, not a constant. A bar for a print proof and a bar for a screen render are different numbers with the same units, and a lib that picks one has decided something it has no standing to decide. The caller declares it and the gate prints it. license_tier: ORIGINAL No hw writes (Rule 26). LIB (no main).