nx_meshmerge.nx

nx_meshmerge.nx -- SPLICE ONE NXMSH2 INTO A NAMED LAYER OF ANOTHER, WITH A PLACEMENT TRANSFORM. ★WHY THIS EXISTS, and it is the measured integration point rather than a guess. `nx_skullsdf` emits a skull that scores headline 73 / detail_head 63 against the 171,248-triangle BodyParts3D oracle, while the LIVE head path (`nx_skullgen`, seven lofted tubes) scores 46 / 14 -- +59% and 4.5x, DARK for weeks. It is dark because nothing could put it INTO the body: `nx_body_gen` consumes a CANON (rules -> rings) and emits three layers, and there was no way to hand it a finished MESH. ★THE AUDIT RAN FIRST (this lane's law) AND IT CHANGED THE DESIGN TWICE. (1) NOT a conversion to the "shared" format: `nx_mesh3` calls itself the ecosystem interchange and is statically capped at 4096 verts / 8192 tris. Our 2mm skull is 48,814 / 97,660 and the oracle is 171,248 tris -- it holds about 5% of ONE skull. The shared format cannot carry this work. (2) NOT a widening of `nx_body_gen`, whose signature is already 24 arguments. Compose, never widen. ⇒ merge INSIDE NXMSH2, in its own organ. ★THE HARD PART IS THE FORMAT'S CONTRACT, NOT THE APPEND. NXMSH2 layers are CONTIGUOUS TRIANGLE RANGES (start,count), so triangles cannot simply be appended -- adding to a middle layer would leave every later layer's range wrong. This regroups ALL triangles by layer and rebuilds the table, which is the same contract `nx_headcrop` already had to honour when it cropped a band. nx_meshmerge <out> <base> <add> <layer> [scale_permil] [tx] [ty] [tz] nx_meshmerge selftest layer = index into the base's layer table (0-based). scale_permil defaults 1000 (identity). license_tier: ORIGINAL expect_exit: 0 No hw writes (Rule 26).