nx_devpaint_lib.nx

nx_devpaint_lib.nx -- THE DERIVED-SCALE DEVIATION HEATMAP (2026-08-26). WHY, MEASURED: /exceed/vrm/seed-san.html has been telling visitors that /exceed/vrm/nishi_seedsan_heatmap.glb is "the INTERACTIVE deviation heatmap (drag-rotate the painted gap itself)". That file was BYTE-IDENTICAL to nishi_fitted_seedsan.glb (sha256 1ef8ec04..., 9,435,856 B each): a copy of the unpainted mesh with no deviation on it anywhere. Rule 25 says rewrite it better rather than filter it out, so this lib is the emitter that makes the claim true. WHAT IS NEW HERE, versus nx_meshdist's incumbent `paint` verb, which is otherwise a working heatmap: 1. THE SCALE IS DERIVED FROM THE MEASURED DISTRIBUTION, NOT PASSED IN. md_paint takes `band_um` as an argument and defaults it to 5000 -- a number the caller has to guess, and the published run guessed 10000. A guessed band is a magic number wearing an argument's clothes: pick it too small and every triangle saturates magenta, too large and the whole body is green, and BOTH pictures look authoritative. Here the ramp ceiling is the measured p95 rounded up to the next whole millimetre, and dp_provenance names that rule so the number travels with the picture. 2. IT MEASURES PER VERTEX, NOT PER CENTROID. md_paint takes the distance from each triangle's centroid. A centroid can sit on the reference surface while a corner is far off it, so the incumbent can hide deviation at exactly the sharp features you built a heatmap to find. 3. THE RAMP IS CONTINUOUS, not five hard bands, so a 39 mm and a 41 mm triangle no longer land in different colours because of where a band edge happened to fall. COMPOSED, NEVER RE-IMPLEMENTED: the distance itself is nx_mmdev_lib's exact point-to-triangle ruler (Ericson 5.1.5 + its uniform grid, integer, no sampling). The colour codec is nx_nxmesh_lib's. This file adds the derivation and the ramp and nothing else -- there is no second distance ruler here. DECLARED IMPRECISIONS, all three: (a) NXMSH2 carries ONE colour per TRIANGLE. Three measured corner deviations are therefore reduced to one colour by MAX -- a triangle is as bad as its worst corner. That is wrong in the direction of SHOWING deviation rather than hiding it, which is the safe direction for an inspection tool. Measured on the shipping subject, a triangle spans about 10 mm on a 1.6 m body and renders about 2 px at showcase size, so per-corner interpolation would change little that a viewer can see; the reduction is a format limit, not a shortcut, and lifting it needs a per-vertex colour source in nx_mesh2glb, which is named here rather than silently wished for. (b) Percentiles come from a histogram of MD_HIST_BINS one-tenth-millimetre bins, so they are exact to the bin. Deviations past the last bin land in it and are counted, so p95 can never be reported BELOW a deviation that fell off the end. (c) The scale is per-subject by construction. Two heatmaps are only comparable by reading the printed bounds, which is exactly why dp_provenance exists and why the CLI prints them. license_tier: ORIGINAL