Summary

• The CCD (GJK/EPA) path inflates each geom by 0.5 * pair_margin in the support() function, but the returned dist was never corrected back to the true surface-to-surface distance. This caused dist to be off by pair_margin compared to the primitive narrowphase, breaking the constraint pipeline (pos = dist - includemargin) which expects un-inflated distances.
• Adds dist += margin after the early-exit check in eval_ccd_write_contact to restore consistency with the primitive narrowphase path.
• Adds regression test test_ccd_margin_dist that verifies CCD contact distances match classic MuJoCo for two boxes separated by 0.05 m with margin=0.1, gap=0.1 on each geom.

Details

The support() function in collision_gjk.py inflates shapes by 0.5 * geom.margin (line 189-190). Both geoms are set to margin = pair_margin before CCD runs, so the total inflation is pair_margin. The GJK/EPA distance is therefore relative to the inflated geometry, not the true surfaces.

The primitive narrowphase (e.g. sphere_sphere, box_box) reports true geometric distances without margin. The constraint pipeline then checks pos = dist - includemargin < 0 to activate constraints. Without this fix, CCD pairs activate constraints too early because their dist is already margin-biased.

Example: Two boxes 0.05 m apart with pair_margin = 0.2:

• Primitive path: dist = 0.05 (correct)
• CCD path before fix: dist = -0.15 (wrong, off by pair_margin)
• CCD path after fix: dist = 0.05 (correct)

Test plan

• New test_ccd_margin_dist passes — CCD dist matches classic MuJoCo
• All 52 existing collision_driver_test tests pass
• Verified test fails without the fix (dist mismatch)