RFD 1139: Scale, up, forward and handedness are measured
Problem
Three defects came from assuming a convention, each producing a result that looked correct. The 2D fit projected components 0 and 1 while the rig is Z-up, so the solver rotated every body ninety degrees and reported 0.022% of stature. stature measured the body’s depth for the same reason, normalising by 0.434 where 1.660 belonged. The camera grid placed cameras at world azimuth while its phrases claim a subject’s front, so 96 frames labelled “front view” were profiles.
None raised anything. A left-right swap is worse: it reprojects onto the same pixels, reports the same residual and passes the referee, because both measure reprojection.
Decision
check_conventions.py measures four properties off the geometry and refuses on a mismatch. render_grid.py runs it before rendering a frame, and reports which checks could not run rather than passing over them. SKILL.md gives the order and the traps.
Forward is the chest, the normal to the shoulder line. Gaze and feet are secondary: on the hv_1 fit the chest and gaze agree to 2.8 degrees while the feet sit 55.3 degrees away, so a stance is not a defect.
Handedness takes three witnesses. The .R against .L split cannot catch a mirror alone, since right comes from the chest and the chest mirrors with the body. Signed mesh volume and basis determinants do.