nx_glbnorm_gate.nx

nx_glbnorm_gate.nx -- proves that every NORMAL nx_gltf_export writes is a UNIT vector per glTF 2.0, and proves it BY DECODING THE ACTUAL binary32 BYTES OUT OF THE BIN CHUNK OF A FILE THE SUBJECT WROTE. WHY THIS GATE HAD TO EXIST. Until 2026-08-25 every .glb this estate published was invalid. The number came from OUTSIDE: KhronosGroup/glTF-Validator 2.0.0-dev.3.10 under node, over all 18 .glb files pulled from https://nishifamily.com by curl, uncapped -- files=18 clean=0 with_errors=18 threw=0 sum=18, and 33,682 errors, of which ACCESSOR_VECTOR3_NON_UNIT x 33,656. SIX of our own gates ran GREEN over that same code the whole time. They were not weak gates; they were gates that measured our code against our own expectations, and the expectation itself was the bug ("renderers re-normalise" -- a comment, never a measurement). This gate exists so the estate can see that class WITHOUT an external referee: it decodes the emitted bytes and applies a tolerance DERIVED FROM THE REFEREE'S OWN SOURCE LINE if (Math.abs(Math.sqrt(r) - 1) > 0.00674) -> ACCESSOR_VECTOR3_NON_UNIT as GN_TOL_FP. A green here is a claim about the SPEC, not about our taste. THE ANTI-VACUITY CONTROL IS THE WHOLE POINT. gg_prefix_f32 below is a VERBATIM TRANSCRIPTION of the emission this fix replaced. T2 shows it FAILS the identical predicate the fixed path passes, so the tooth can tell a fix from a no-op. T2b proves the transcription is faithful by sweeping it against the still-live gl_i2f32 rather than trusting that it was copied correctly. WHY THIS FILE IS ALL SMALL FUNCTIONS AND main() HOLDS ALMOST NOTHING. NOT STYLE -- A MEASURED LANDMINE. The first draft did the whole job inline in main(), which reached roughly 85 local bindings, and the compiler then MISCOMPILED SOME OF THEM WITH NO DIAGNOSTIC AT ALL. Measured, from that build's own output, with the constants and the inline product printed beside the binding: DUMP GG_NV = 145 ... GG_VEC3_F32_BYTES = 12 ... GG_NV * GG_VEC3_F32_BYTES inline = 1740 DUMP gg_pb = 1756 <- let gg_pb: i64 = GG_NV * GG_VEC3_F32_BYTES DUMP gg_nb = 5244 <- let gg_nb: i64 = GG_NV * GG_VEC3_F32_BYTES (the SAME expression) Two `let`s of one constant expression, in one function, disagreeing with each other AND with that expression evaluated inline one line away. Everything computed from a RUNTIME value in the same function (jlen 724, binoff 752, binlen 6384, fsz 7136) stayed correct and self-consistent throughout. The cost of that landmine was a full RED that looked exactly like a broken fix: the gate read the normal block 8 bytes late and reported 132 of 145 normals non-unit, while the writer had been correct all along -- the real block sat at 752+1740, exactly where it belongs. ***A MISCOMPILED LOCAL IS A SILENT WRONG ANSWER THAT INDICTS THE SUBJECT INSTEAD OF THE INSTRUMENT.*** So every function here stays small, and the BIN layout is derived ONCE, in gg_layout, from PARAMETERS rather than from constant-expression locals. THE GATE READS u32s WITH ITS OWN gg_ru32 AND ENCODES EXPECTED POSITIONS WITH vm_int_to_f32 RATHER THAN THE SUBJECT'S gl_i2f32 -- ON PURPOSE, and for the reason nx_vnormals_gate states: an oracle that borrows its subject's helpers agrees with it by construction and stops being able to detect a change in them. The side effect is a free measurement: T8 passing means gl_i2f32 and vm_int_to_f32(v,1) are the same