nx_fieldfit_verified_r3_t333.nx

nx_fieldfit_lib.nx -- FIT A PARAMETRIC IMPLICIT FIELD TO A SCANNED ORACLE. The machinery nx_skullsdf's `tune` grew (anatomy AN17, 2026-09-18/19), extracted so any anatomy field -- a skull, an arm, a foot, a muscle belly, a tissue layer -- is fitted by ONE optimiser with ONE objective and ONE fit-file contract, and a defect fixed here is fixed for every field at once. The consumer hands over three functions as VALUES (NishiLang func-pointer locals and fields, indirect calls up to six arguments): field(px, py, pz, a, b, P) -> signed distance at `fq` units per millimetre (negative inside) for parameter vector P, with a and b the consumer's two policy knobs (the skull passes sex and robusticity) region(x, y, z) -> 0 .. nreg-1, the named anatomical region a point belongs to live(k) -> 1 for a station the field reads, 0 for a retired one (the descent skips it: two evaluations spent on a dead station are a pass that cannot move anything) THE OBJECTIVE IS TWO-SIDED, AND IT HAS TO BE. A one-directional score -- how far is each ORACLE point from my surface -- is trivially gamed by DELETING GEOMETRY: the skull's first tuning run shrank its orbits from 22 mm to 8 and its maxilla from 30 mm to 13, improving the number 40 percent by filling in the eye sockets, because surface that does not exist cannot be wrong. The REVERSE term charges every lattice point on OUR zero level its distance to the nearest oracle sample, so a deleted feature is EXPENSIVE by construction. (The reverse term uses the nearest oracle VERTEX from a spatial hash: an upper bound on the true point-to-surface distance -- it can overstate, never understate, the safe direction for a penalty.) THE OBJECTIVE IS REGION-BALANCED. A plain mean is dominated by whatever region has the most samples (the skull's vault and base held 9,335 + 8,672 of ~31,000, so a 49 mm muzzle and 4 mm drilled orbits read as a 4 mm skull and the surface objective fell 629 -> 531 while the chin moved +4 -> +19.6 mm). Each side is the mean of the PER-REGION means (a region with no samples abstains), so the smallest region weighs what the largest does. No weight is typed: the population decides. THE DESCENT is coordinate descent: per live station try +step then -step, keep a move only if the MEASURED objective falls; the step halves per pass, floor 1. No change is ever kept on the strength of looking better. THE FIT FILE is newline-separated integers, loaded as a PREFIX: a file shorter than the vector applies to its first entries and the rest keep their defaults, so growing a parametrisation never discards banked convergence; a file shorter than `minlen` changes nothing (parsed into a temp and applied only when long enough, so a truncated write cannot half-poison a field). Written whole by ff_save. UNITS: objectives are reported in centi-millimetres; `lat` (the reverse lattice spacing), `ext` (the half-extent the hash covers) and `dim` (cells per axis) are the consumer's, taken from the resource it fits. LIB, no main. license_tier: ORIGINAL. No hw writes (Rule 26).