nx_nxa_write_lib.nx

nx_nxa_write_lib.nx -- THE NXANIM01 CONTAINER WRITER AS A LIBRARY (/compare/modding MD2, 2026-09-06). The estate held three INLINE emitters of the same container (nx_meshrig, nx_fbx2nxa, nx_gltf2mesh) and one append-only section writer (nx_nxa_posewrite_lib); a fourth reader (XPS) would have been a fourth copy. This lib composes nx_nxa.nx -- the ONLY place the magic, the tag encoding and the rolling checksum are defined -- so a writer and every reader cannot drift. LAYOUT (knowledge/nxa_format_spec.md, as nx_mesh2glb and nx_meshrig read it): header words [magic][version=1][nsect] [toc check]; TOC entries of 4 words [tag][byte_off][word_len][check]; payloads of i64 words. VERT [nv][x y z ...] in 0.01 mm model units; TRIS [nt][a b c ...] one word per index; SKEL [nj] + 8 words per joint [parent][tx ty tz] [qx qy qz qw q12] with the bind rotation the IDENTITY quaternion by the estate's rig convention (stated, never smuggled); SKIN [nv] + 8 words per vertex [j0 j1 j2 j3][w0 w1 w2 w3] with the four q12 weights summing to exactly 4096. EVERY REFUSAL IS NAMED: a triangle index outside the vertex table, a skin joint outside the skeleton, a vertex whose weights do not sum to 4096, or an empty mesh, all refuse BEFORE a byte is written -- the container never carries a section the spec would call malformed. SKEL and SKIN are written together or not at all (nj == 0 writes VERT and TRIS only). license_tier: ORIGINAL No hw writes (Rule 26).