Confirm the absent grasp at solver level, and name the inverted finger command

The published grasp finding inferred "the pads never touch the handle" from
finger link origins, which is weak: a link origin can sit far from the pad
surface it carries. Replaced with a measurement that cannot be argued with. A
contact sensor filtered to the two inner-finger bodies reports exactly 0.0 N
against the blade at every commanded closure across the full finger_joint range
of 0 to 0.8203 rad, and the joint reaches every commanded value exactly with
zero drive torque, so nothing obstructs the pads anywhere in their travel.

The same sweep upgrades the secondary observation to a second defect. The joint
limits are 0 to 0.8203 and pad separation is 0 mm at 0 rad, so zero is the
closed end and separation grows with the command. The task calls 0.45 pregrasp
and 0.60 closed, so the command it treats as closing opens the fingers by a
further 16 mm. That is now a result rather than a lead, and it will surface the
moment the frame error is corrected.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: tryaksh

docs/status.md

@@ -344,10 +344,14 @@ force grid from 0 to 120 N, so 128 environments sweep the surface in one run.
 
 Report: `evidence/grasp_axial_pull_gate.json`.
 
-**There is no grasp to characterise.** The finger pads settle 165 mm away from
-the handle and never touch it, the drive joints reach their commanded closure
-without developing torque, and the blade is a free-floating body in zero
-gravity. Its motion under load confirms this arithmetically rather than by
+**There is no grasp to characterise.** The solver says so directly: with a
+contact sensor filtered to the two inner-finger bodies, PhysX reports exactly
+0.0 N between pad and blade at every commanded closure across the full
+`finger_joint` range of 0 to 0.8203 rad, and the joint reaches every commanded
+value exactly with zero drive torque, so nothing obstructs the pads anywhere in
+their travel. That measurement does not depend on where the pad surfaces are
+assumed to be. The pads settle 165 mm away from the handle and never touch it,
+and the blade is a free-floating body in zero gravity. Its motion under load confirms this arithmetically rather than by
 inspection: 120 N on a 10 kg blade for 1.5 s predicts 13.5 m of free travel, and
 the measurement is 12.5 m. The earlier "failed pull gate" was not a weak grip.
 
@@ -375,12 +379,12 @@ blade. Three consequences:
   three certifications. The grasp task needs its own corrected offset, leaving
   the insertion tasks on the value they were trained against.
 
-A secondary observation needs confirming before it is acted on: finger-pad body
-separation grows monotonically with the commanded value, from 0 mm at 0.00 rad
-through 42.7 mm at 0.45 to 58.9 mm at 0.60. The task calls 0.45 "pregrasp" and
-0.60 "closed", so the command it treats as closed separates the pad bodies
-further than the one it treats as open. That was measured between pad *body
-origins*, not certified pad faces, so it is a lead rather than a result.
+A second defect sits behind the first and will bite the moment it is fixed: the
+finger command runs backwards. Pad separation grows monotonically with the
+commanded value, from 0 mm at 0.00 rad through 42.7 mm at 0.45 to 58.9 mm at
+0.60, and the joint's own limits are 0 to 0.8203 rad. Zero is therefore the
+closed end. The task calls 0.45 "pregrasp" and 0.60 "closed", so what it treats
+as closing the gripper opens it by 16 mm.
 
 **What this does and does not invalidate.** It does not touch the promoted
 Level-0/1/2 insertion results. Those depend on blade-versus-slot geometry, rail