feat(lite_biped): re-export feet from CAD, model the full shoe sole

[T-K-233]

Why

The shoe sat in the wrong place in the previous CAD export. Re-exported from Onshape (same document / configuration) offline via robot.pkl — no API traffic.

The kinematic change is confined to the two foot links: COM and visual origin move ~3.8 mm, inertia rises 13–16%, mass unchanged. Standing height is not affected — the sole plane stays at exactly −0.775 m, so the training init pose, foot sites and every height-dependent reward are untouched.

Foot colliders: 2 boxes → 11 spheres per foot

The old boxes only modelled the flat patch. The shoe bottom actually continues into a 21 mm toe upturn over y = 58 → 132 mm that no collider reached:

tilt from horizontal	area	height above sole	y range
0–2° (flat)	90.4 cm²	0.00–0.25 mm	−73.8 … 60.9
2–10°	24.1 cm²	0.20–2.52 mm	57.9 … 84.0
10–25°	28.6 cm²	2.68–12.93 mm	84.9 … 118.1
25–45° (toe tip)	6.5 cm²	12.46–20.73 mm	117.1 … 132.4

Sphere centres sit on the 3D lower convex hull of the sole — the only points that can ever touch a plane, since everything else is shadowed for every possible ground orientation — and are offset along the inward surface normal, so each is tangent to the shoe rather than hovering above it (on a 25° toe facet those differ by 3.4 mm).

Validation

Swept pitch −20…+40° × roll ±10°, comparing the sphere set's lowest point against the real mesh at each pose — exactly the quantity contact resolution sees:

clearance error   mean +0.46   max +2.07   min -0.95 mm
standing support  97.96 cm² across 8 load-bearing spheres

Spheres are also cheaper than the boxes they replace: box-plane emits a contact per bottom corner, so grounded boxes cost up to 16 contacts per foot versus 11 here, with no near-duplicate corner pairs for the solver to arbitrate.

Also folded in

Earlier work on this branch that the regenerated artifacts interleave — splitting it would mean committing a generator that disagrees with its own output:

• MJCF actuators switch from <motor> to <position>, putting the joint PD inside the physics step instead of sampling it once per controller_manager tick (regenerates lite_bimanual and lite_dummy too). Its unit test still asserted <motor> and was failing — updated to match.
• ankle_pitch four-bar linkage parameters in cad/ros2_control.json and the xacro generator.

Test plan

• uv run pytest — 1418 passed, 5 skipped
• MJCF compiles in MuJoCo: nq=21, nv=20, ngeom=41; 11 collision spheres per foot carrying condim=3 friction="0.6 0.005 0.0001" priority=1
• foot_collision_N geom naming preserved, so downstream friction DR keeps matching

🤖 Generated with Claude Code

https://claude.ai/code/session_012QKJQ9hze1hMWXJpdPcjp3