fix(box_task): re-fit cardboard_box visual walls to its frozen collision hull
The scanned visual mesh under-reached the 5-primitive collision hull: below the flap
flare the walls sat at |x|<=0.0977 / |y|<=0.1016 while the hull inner faces are at
0.108 / 0.1105 — contents resting against the physics wall rendered poking ~8.3 mm
through the cardboard (measured: a settled soda can's outer edge at rel y +0.1085 vs
the old visual sheet at 0.1002).
Fix: monotone piecewise-linear radial remap of box.obj + usd/cardboard_box.usd
(x knots 0->0, 0.070->0.070, 0.0923->0.1083, 0.1254->0.1230; y knots 0->0,
0.075->0.075, 0.1002->0.1108, 0.1255->0.1255, world frame). Wall interior surfaces now
sit 0.3 mm outside the hull inner faces; flap tips end exactly ON the hull outer faces.
UVs/faces/topology untouched; OBJ and USD agree to 2e-7 m.
ZERO physics change: cardboard_box.xml geoms are byte-identical (only the doc comment
grew); a before/after settle probe through hand_track.pack_box left the can's resting
pose bit-identical. RoboVerse spawn-clearance lock tests
(roboverse_pack/tasks/hand_track/tests/test_fruit_tray_spawn_clearance.py) pass 5/5.
Spawn placement must still clear the HULL, not the render: an invisible ~1.2 cm outer
margin below the flap flare remains by design (the hull is frozen — shared by 7 tasks,
aligned with Parrot's recorded contact dynamics, checker-calibrated).