nx_faceanat_scale_gate.nx

nx_faceanat_scale_gate.nx -- THE REFEREE FOR THE FACE MODEL'S SCALE (aesthetictwin AT17, 2026-09-02). THE DEFECT: nx_faceanat declared MMU = 35 (3.5 units per mm) from a comment that took a 150 mm HALF-width for a skull that is about 145 mm WIDE, and every millimetre the aesthetictwin board publishes, and every tissue relief the face emitter draws, went through that constant. Four witnesses of the model's OWN declared landmarks against the ANSUR II female means say about 7 to 8.7 units per mm -- a 2.2x error that the board carried as an open contradiction since 2026-08-28. THE MECHANISM: knowledge/faceanat_scale.conf carries one WITNESS row per landmark pair (the model's declared span in units, the ANSUR II female mean in tenths of a mm, n, the refs key), one DERIVED row (the median units-per-mm x10 with the min and max of the witnesses) and one SPREAD row (the widest witness disagreement tolerated, permil of the median, with its reasoning beside it in the conf). This gate RECOMPUTES the derivation from the witness rows and holds nx_faceanat's fa_mmu_x10() to it -- a typed constant that drifts from its own derivation cannot pass -- and its neg-control asserts the LEGACY 35 lies OUTSIDE the witness band: the tooth that could not have passed before this rung existed. The population bound the reader prints (a military sample, not a civilian approximation) travels with the number in the conf and in the verdict. THE SPREAD IS A SEPARATE FINDING FROM THE SCALE: the four witnesses disagree by 218 permil (MEASURED live 2026-09-02), which is the model's PROPORTION error landmark to landmark, not its scale. The bar for it is a conf row so it can be RATCHETED DOWN as the proportions are re-fit -- never a literal here that gets loosened to flatter the number. SCOPE: this proves the SCALE. It does not re-baseline the emitted relief -- that is a visual acceptance under the operator's three-party rule and is a separate rung. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0