nx_partloft_lib.nx

nx_partloft_lib.nx -- THE LOFTED PART: Infinigen's NurbsPart, replicated in NishiLang from the first byte up (procgen PG59 / anatomy AN19, 2026-09-19). Operator standing order the same day: the estate uses the best-in-class published method as its BASELINE and improves it piece by piece; the ellipsoid-union skull is retired. This file is the first piece. ONE part = a skeleton polyline with a normalised profile at every station, DECOMPOSED into the genes the Infinigen creature system mutates, RECOMPOSED into a grid of NURBS control handles, EVALUATED as a cubic B-spline surface, and WRITTEN as NXMSH2 through the estate's one mesh layout (nx_nxmesh_lib). Read from the primary source, mirrored under knowledge/fetched/ and pinned on procgen.refs: infinigen/assets/creatures/parts/generic_nurbs.py decompose_nurbs_handles / recompose_nurbs_handles / NurbsPart.sample_params (the gene noise STRUCTURE) infinigen/assets/creatures/util/geometry/lofting.py factorize_nurbs_handles, skeleton_to_tangents, compute_profile_verts (frames by rotate_match_directions) infinigen/assets/creatures/util/geometry/nurbs.py geomdl_nurbs: degree (3,3), clamped uniform u knots, cyclic v wrapping three control points, sample_size from face_size infinigen/assets/creatures/util/part_util.py nurbs_to_part: the attachment skeleton is the INTERIOR rows infinigen/assets/creatures/util/creature.py apply_attach_transform: (u, v, rad) surface coordinates CONVENTIONS, THEIRS: a part runs along +x (profiles face +x), y is lateral, z is up; skeleton_yoffs is the lateral drift per segment and thetas the pitch in the xz plane (degrees there, VM_ONE radians here). UNITS, OURS, and there are exactly three: POSITIONS in tenths of a millimetre (u10, NM_UNIT_PER_M, the NXMSH2 ruler's unit); FRACTIONS in PL_FQ = 65536 (Q16) -- gene proportions, lateral drifts, station parameters, normalised profiles, surface parameters, the attachment coordinates, the gene variance, and the B-spline knots and basis; and ANGLES in VM_ONE radians (Q12), the estate's one trigonometric unit, for the pitch genes and the yaw. A radius is a position with sixteen fractional bits (Q16 u10), because a profile normalised by a radius rounded to 0.1 mm was measured 0.25 percent off at 20 mm and would be 2.5 percent off on a finger: MEASURED on the first run of the gate, where Q12 proportions truncated to whole u10 per segment lost 0.5 percent of a part's length (995 for 1000). Every division that lands on a stored value rounds to nearest (pl_rdiv); truncation is how that half-percent was lost. No float anywhere: the file's float32 is encoded once, at the door, by nx_nxmesh_lib / nx_vecmath. DECLARED DEVIATIONS (each a rung on procgen, never a silent difference): 1. np.random.normal is Irwin-Hall (twelve uniforms of a 31-bit LCG): same mean and variance, tails end at six sigma. The STRUCTURE of the noise is the source's exactly: one size draw, one length draw, one shared radius draw plus one per station, one per proportion, zero-meaned pitch noise, per-column times per-cell profile noise mirrored across the profile. 2. The attachment ray-cast against a BVH is a parametric search: around the requested station the closed profile is searched for the point nearest the ray (least perpendicular distance, in front of the origin), coarse then bisected, over a one-segment window of u. 3. The attachment skeleton is the clamped cubic B-spline of the interior rows, where the source runs two Catmull-Clark levels over the interior polyline (the limit of that subdivision IS the cubic B-spline). 4. The emitted mesh carries two fan caps so the part is a closed solid for the lattice fusion; the source leaves the tube open and lets remesh close it. Winding is made outward per triangle here; the source runs normals_make_consistent. LIB, no main. license_tier: ORIGINAL (a re-implementation from a read of BSD-3 source; no source bytes copied; the