nx_gsplat_fit_aniso_gate.nx

nx_gsplat_fit_aniso_gate.nx -- FITTING ON THE QUALITY PATH: the anisotropic surfel reconstructor. WHY THIS GATE EXISTS. nx_gsplat_fit_gate proved the ISOTROPIC path can be fitted from an image. That is the blob path. The ANISOTROPIC path is the one that renders oriented surfels -- the representation that carries actual quality -- and until 2026-08-23 it could be rendered but never FITTED. A reconstructor that stops where the representation starts to matter is not a reconstructor. ★★THE ACCEPT RULE IS THE ONE THE ISOTROPIC GATE'S FIRST RED TAUGHT US, AND IT IS INHERITED DELIBERATELY. v1 of that gate demanded the error fall at EVERY step; the optimiser reached loss=0 and correctly STOPPED, so the rule demanded the impossible and would have REWARDED A JITTERER. The rule splits at the only boundary in the mathematics: WHILE loss > 0 -- a residual remains, so descent must STRICTLY reduce the error. ONCE loss == 0 -- converged, so the error must be EXACTLY CONSTANT. Both phases must occur or the tooth is vacuous and says so, and the convergence floor is DERIVED: an image cannot localise a gaussian finer than a pixel, so world-units-per-pixel IS the floor. ★T0 IS THE TOOTH THAT MAKES THIS GATE DIFFERENT FROM ITS ISOTROPIC SIBLING. If the fixture's surfels happened to project to CIRCLES, the anisotropic gradient would reduce term-for-term to the isotropic one and this gate would prove NOTHING NEW while passing everything. So the conic is measured and the scene must be genuinely elliptical -- ca != cc or cb != 0 -- before any other tooth is believed. license_tier: ORIGINAL expect_exit: 0