nx_fitloop.nx

nx_fitloop.nx -- THE KEYSTONE: METROLOGY IN THE AUTHORING LOOP (debt 1785967773; operator 2026-08-05: crafting and measuring must grow each other). A generator is driven by a judge in a CLOSED LOOP with an auditable trial trace -- the capability this session proved manually (hand 241->358->258->191->181mm across five hand-run pose trials) and the gap measured in the industry: authoring tools have no metrology in the loop, metrology tools do not author. v1 searches the 2-parameter pose space measured tonight: symmetric shoulder raise s and elbow raise e (degrees, clamped 0..90; left rows get -deg, right +deg; elbow pivot approximation declared valid for small s). Generator = nx_meshpose apply. Judge = nx_meshdist parts; OBJECTIVE = the SUM of all per-part mean distances (pose moves only arms, torso terms are constant, so minimising the sum minimises the arm gap without wrecking anything else -- and a whole-body objective cannot be gamed by improving one part at another's expense). u2605THE VACUOUS-OBJECTIVE GUARD IS THE ORGAN'S FIRST TOOTH: a failed judge run yields an empty capture whose sum is 0 = "perfect". Any trial parsing fewer than 10 part rows is INVALID and scores WORST, never best -- an optimiser that can reward a broken ruler will find the broken ruler immediately. nx_fitloop pose2 <base.nxmesh> <oracle_parts.nxmesh> <out_best.nxmesh> [budget] nx_fitloop selftest Trial trace: one TRIAL row per evaluation (s, e, objective um-sum, parts parsed) -- the optimisation is EVIDENCE, not a black box. license_tier: ORIGINAL expect_exit: 0