nx_meshrig_gate.nx

nx_meshrig_gate.nx -- IS THE BIND REAL, OR JUST A SKIN SECTION THAT EXISTS? Subject: ./nx_meshrig.elf (the serving-root binary; e2e fork, never an in-process re-derive). THE DEFECT THIS GATE EXISTS TO CATCH: a RIGID bind -- every vertex nailed to one bone at full weight -- satisfies "the asset has a SKIN section", loads in nx_nxa_play, reports zero_weight_verts=0, and produces candy-wrapper garbage the moment anything poses. It is the trivial wrong implementation, and presence-of-section cannot distinguish it from a real bind. Two numbers can: under a rigid bind EVERY vertex has exactly one influence and the mean dominant weight is 4096 (the q12 whole). THE NEG-CONTROL NEEDS NO CODE MUTATION, WHICH IS WHY IT IS TRUSTWORTHY: a skeleton with ONE bone IS a rigid bind, by construction. So the rigid case runs the SAME promoted binary down the SAME code path with a different fixture -- there is no mutant to restore and no risk of measuring a harness instead of the subject. A 4-joint skeleton must blend; a 2-joint skeleton must not. FIXTURES ARE ASSEMBLED AT RUNTIME under /tmp/nxmeshriggate/ (a gate must not share its fixture with a production beat, and a teardown does not run when a run crashes). The mesh fixture is emitted here as real NXMSH2 bytes, including a real IEEE754 f32 encoder, because a fixture the defect cannot fail is not a test. SUBPROCESS CAPTURE COMPOSES nx_gatekit_lib. This is not a style preference: a hand-rolled pipe/fork/exec in a sibling gate DEADLOCKED IN PRODUCTION today (50+ minutes in pipe_wait, zero CPU, artifact frozen at its header) because an undrained pipe blocks forever. The banked law is that fixtures and subprocess come from that lib; never re-roll them. license_tier: ORIGINAL expect_exit: 0