nx_recon3d.nx

nx_recon3d.nx -- MULTI-VIEW TRIANGULATION, the classical MVS core of photo->3D reconstruction (cadtwin P1, the "photograph -> 3D" capture half). This is the CLASSICAL / no-trained-model path (Hartley-Zisserman multi-view geometry / COLMAP triangulation) -- the part a SOVEREIGN no-float stack CAN build; the 2025-26 feed-forward SOTA (DUSt3R/MASt3R/VGGT) is neural = trained-model-bound (our named gap). Benchmarked by CHAMFER DISTANCE (the DTU metric the field scores reconstruction on): synthesize N camera views of a KNOWN point cloud -> triangulate each point back -> integer Chamfer vs ground truth. Deterministic, no dataset dependency, no float. Composes toward the full pipeline (feature match nx_features + pose est = later rungs). license_tier: ORIGINAL