nx_editjudge_gate.nx

nx_editjudge_gate.nx -- proves the edit-compliance judge distinguishes the three outcomes that a single "did it change" number would blur, and that its hue primitive is correct. Imports the same lib the tool ships. T1 PERFECT SHIFT : ref yellow, gen green -> compliance ~1000, shifted == masked. T2 NO CHANGE : ref yellow, gen STILL yellow -> compliance ~0, stayed == masked. THIS IS THE EXACT FAILURE the retention loop hid; the judge MUST score it near zero. T3 WRONG CHANGE : ref yellow, gen blue -> compliance ~0, drifted == masked. "changed" but not to target must NOT count as compliance -- a magnitude-only judge would pass this. T4 PARTIAL : half the masked pixels shifted -> compliance ~500 (orders correctly). T5 EMPTY MASK : ref all gray (no from_hue garment) -> masked 0, compliance 0, NO divide fault. An edit with nothing to edit is not a passing edit. T6 HUE MATH : pure R/G/B/Y map to 0/120/240/60 degrees (the primitive is correct, not lucky). T7 SAT GATE : a DESATURATED pixel at from_hue is excluded from the mask (a washed-out near-gray cannot be trusted to carry a hue, and would poison the count). T8 ORTHOGONALITY : the identical image (ref==gen) scores compliance 0 -- i.e. maximum RETENTION yields minimum COMPLIANCE, which is the whole reason a second judge exists. expect_exit: 0 license_tier: ORIGINAL