nx_nxa_retarget_gate.nx

nx_nxa_retarget_gate.nx -- THE DONOR RETARGET GATE: handedness is DERIVED, limb length is preserved BY CONSTRUCTION, and a witness that cannot testify ABSTAINS BY NAME. WHY ITS NEGATIVE CONTROLS ARE STRONGER THAN ANY MUTANT I COULD CONSTRUCT: every RED below is a REAL DEFECT THIS ORGAN ACTUALLY SHIPPED AND THIS LANE ACTUALLY CAUGHT, on the real corpus, within hours. Nothing here is synthetic. - CHIRALITY. Before the mesh-derived anterior cue, frame_det came out +1 / +1 / -1 / -1 across four ordinary humanoids -- i.e. TWO OF FOUR DONORS WERE POSED MIRRORED, left and right swapped, and NO LIMB-LENGTH TOOTH COULD CATCH IT because a mirror preserves every length it checks. The cause: handedness was resolved by argmin over mapping distance, and a bilaterally symmetric point set (every rig here scores sym_permil=1000) is near-ISOMETRIC UNDER MIRRORING, so both options tie and the argmin picks arbitrarily. - VACUOUS BANDS. The anterior cue takes two witnesses, a foot band and a head band, and requires them to agree. With the band height derived from GLOBAL mean joint spacing, paladin measured foot=head=2184 and dark_knight foot=head=-140 -- BYTE-IDENTICAL, because both bands covered the entire mesh. TWO WITNESSES THAT ARE SECRETLY THE SAME WITNESS AGREE ONE HUNDRED PERCENT OF THE TIME. dark_knight's whole verdict rested on -140 units of whole-mesh noise. - UNPOPULATED BANDS. Disjointness turned out to be NECESSARY BUT NOT SUFFICIENT: toon3d8's toe-tip bone is 1 unit long, so its foot band held exactly 2 VERTICES and an entire handedness verdict rested on them. A band must also be POPULATED, and the floor comes from the asset -- a region holding fewer points than the rig has joints is stray vertices, not a shape. The three donors that now REFUSE are therefore the gate's own neg-controls: they prove the teeth can fail, on real assets, for three DISTINCT named reasons. usage: nx_nxa_retarget_gate (no args -- reads the real corpus) license_tier: ORIGINAL expect_exit: 0