nx_svmetrology.nx

nx_svmetrology.nx -- THE MULTI-CUE SCALE SOLVER (debts 1785958292 + 1785957867; operator 2026-08-05: "be clever but stay reproducible"). Single-view metrology done as an OVER-DETERMINED system. THE DOCTRINE, IN CODE: never choose a reference. Every scale cue in a scene is an independent estimate of ONE unknown (camera height, in micrometres), so N cues leave N-1 degrees of freedom of VALIDATION. Fuse by inverse-variance weighting, then TEST the agreement with chi-square. Below threshold: report the interval. Above: REFUSE and NAME THE OUTLIER -- averaging away a disagreement hides the broken assumption that caused it (a partition is a claim; check the parts sum). A SINGLE CUE IS NEVER VALIDATED: 0 dof means there is nothing to check it against, and the organ says so out loud rather than reporting a confident number. GEOMETRY (Criminisi/Reid/Zisserman single-view metrology): for a flat ground plane the horizon sits at camera-optical-axis height, so for any object standing ON that plane, with no focal length needed: h_obj / h_cam = (y_base - y_top) / (y_base - y_horizon) LEVERAGE LAW (measured 2026-08-05): the error scales INVERSELY with (y_base - y_horizon), so a NEAR reference beats a distant one however crisp the distant one looks. The organ reports that denominator as `leverage_px` on every cue it derives, so a caller can see which reference is worth having. nx_svmetrology camheight <y_base> <y_top> <y_horizon> <h_obj_um> <sigma_permil> -> one cue row nx_svmetrology height <y_base> <y_top> <y_horizon> <h_cam_um> <sigma_permil> -> a height, given scale nx_svmetrology fuse <cuefile> [max_reduced_chi2_x100] -> posterior + chi2 + residuals nx_svmetrology selftest CUE ROWS (tab-separated; the cue registry plane uses the same shape): cue <TAB> <name> <TAB> <value_um> <TAB> <sigma_um> [<TAB> provenance [<TAB> source]] license_tier: ORIGINAL expect_exit: 0