nx_mmdev_lib.nx

nx_mmdev_lib.nx -- THE MILLIMETRE SURFACE-DEVIATION RULER (PG4, procgen.plan, 2026-08-24). WHY: every realism claim on the human lane is graded by nx_bodybench in PERMIL of silhouette and normal agreement. A cosmetic surgeon buys none of that. The clinical unit is millimetres of predicted-vs-actual surface deviation, and the published soft-tissue prediction bar sits near 1-2 mm mean error. Without a ruler in that unit every fidelity claim after this is aesthetic opinion wearing a number. WHAT IT MEASURES: for EVERY vertex of mesh A, the exact Euclidean distance to the nearest point on the SURFACE of mesh B (point-to-triangle, Ericson 5.1.5, all integer), then mean, p95 and max. Full population, never a sample: `verts` is printed and the gate binds every assertion to it. UNIT DERIVED, NOT PICKED: coordinates are held in tenths of a millimetre. The barycentric test forms products of two dot products; each dot is bounded by 3 * extent^2. For an extent of 2.5 m at 0.1 mm that is 3 * 25000^2 = 1.9e9 per dot and 3.5e18 per product, under the 9.2e18 i64 ceiling. A finer unit overflows a whole-body subject; a coarser one is too blunt for a 1 mm bar. The extent bound is CHECKED at load and the ruler REFUSES by name above it -- wrong in the direction of doing nothing. ACCELERATION WITHOUT SAMPLING: B's triangles go into a uniform grid whose resolution is derived from the triangle count (about four triangles per cell). A query walks expanding rings of cells and stops when the ring's inner face is provably farther than the best hit so far. Exact, never approximate. DECLARED IMPRECISIONS: (1) A's vertices are visited per triangle-vertex, so a vertex shared by k faces weighs k times -- the mean is valence-weighted, which favours dense regions; printed as `tri_verts`. (2) p95 is read from a histogram with 0.1 mm bins, so it is exact to the bin. (3) One-sided A->B; the gate runs both directions and reports both. license_tier: ORIGINAL