nx_clipcheck.nx

nx_clipcheck.nx -- THE ANIMATION CLIPPING RULER (debt 1786578438; operator 2026-08-12: evaluate art/realism/quality/CLIPPING). Measures, per ANIM key of a chn-2 NXA clip, how deep tracked limb joints penetrate (a) the TORSO CAPSULE and (b) the FLOOR plane -- the two gross clipping classes the eye reports as "arm through chest" and "foot through ground". Joint world position per key = bind + dt (the wire's dt lanes ARE world displacement, mm -> 0.01mm x100), exactly the player's apply contract -- no quat math needed for JOINT CENTERS. DECLARED IMPRECISION (imprecision named or the next reader trusts it as exact): - torso = ONE vertical capsule at the spine-band centroid, radius CC_TORSO_R (data below), span 40..80 permil of stature; limbs = points with radius CC_LIMB_R. Real flesh is neither. - ELIGIBILITY: only joints whose BIND already sits OUTSIDE torso_r+limb_r+grace are measured (wrists/elbows/fingertips/ankles) -- hips/shoulders bind INSIDE the capsule and would be permanent false positives (the bind pose is the clean control by construction). - v1 measures limb-vs-TORSO and vs-FLOOR only, NOT limb-vs-limb. THE BVH LANE (2026-08-13, debt 1786598036) -- the MOCAP-SELF-CLEARANCE CONTROL, the two-instruments law applied to the retarget: the SAME capsule measure over the mocap SOURCE skeleton (FK via nx_bvhfk_lib -- the proven incumbent, never a re-implementation). If the source clips the same capsule at the same keys, the retarget is faithful and the residual is ruler coarseness; if the source is clean, the gap belongs to the emitter. DECLARED: - the reference NXA supplies the STATURE BRIDGE (BVH declares no unit; source positions are scaled so both skeletons share a stature) -- proportions are the subject, not absolute size. - axes are MEASURED from the rest pose: up = bone-length-weighted alignment (a T-pose span exceeds its height, so extent misdetects up); side = widest remaining extent; ambiguous alignment REFUSES rather than guessing gravity. - root POSITIONAL channels are ZEROED per frame: the nxa emitter lane is rotations-only / in-place BY DESIGN, so the control mirrors its subject. Rest pose (all channels zero) is the bind analog for eligibility. Keys were emitted 1:1 from frames, so key == frame. - measures every parser-exposed joint whose rest pose sits outside capsule+grace (the NXA lane measures tracked joints; both predicates are "rest/bind outside"). usage: nx_clipcheck <clip.nxa> [detail <key> <joint>] | bvh <clip.bvh> <ref.nxa> [up=x|y|z] exit: 0 CLEAN (worst <= CC_TOL_MM) | 1 CLIP (worst key+joint NAMED) | 4 refuse (parse) license_tier: ORIGINAL No hw writes (Rule 26).