[paulrobello]

Fixes #1070.

Contacts are found once per step, and each substep re-measures the separation from a body-fixed anchor. For a ball, the true contact point stays at centre + r·n while the ball spins. The body-fixed anchor rotates away and lifts by r·(1 − cos θ). The solver reads that lift as separation, so rolling balls sink until the bias cancels it.

Change

• Adds ContactManifoldData::centered_ball: [bool; 2] (crate-private, serde(default)). It is set in pair_update.rs when a side's collider is a ball centred on its body's CoM.
• For such a side, the twist and Coulomb kernels track the normal separation by translation from the body centre. This covers build time, substep update and update_rhs_wo_bias.
• Friction drift keeps the body-fixed anchor.
• Non-ball shapes, off-centre balls, capsules and the generic (multibody) constraint path are unchanged.
• Cost: a few lane selects per contact point.

Results

Measured on a 16 mm ball at 240 Hz on a flat trimesh. A frictionless sliding ball sits at the soft-contact equilibrium, 0.069 mm, both before and after.

Speed	Penetration before (mean)	Penetration after	Max end-of-tick |vy| before	Max end-of-tick |vy| after
0.3 m/s	0.091 mm	0.069 mm	6.52 mm/s	0.00 mm/s
1.17 m/s	0.641 mm	0.069 mm	20.44 mm/s	0.00 mm/s
2.0 m/s	0.736 mm	0.069 mm	20.44 mm/s	0.00 mm/s

Tests

• New test: pipeline::physics_pipeline::test::rolling_ball_does_not_sink_into_the_ground. It covers both friction models and both collider orders. On v0.36.1 it fails for the Simplified model with the ball first: rolling penetration is 0.00245 against 0.00160 sliding.

• Goldens re-minted, because contact behaviour changes for every rolling ball:

  Test	Old	New
  parallel_path_parity	0xcc9e_c93f_2c15_65cc	0x0332_9d78_889b_bd8c
  simd_backend_determinism	0xe488_2a11_2d57_d212	0x9943_2aa0_b8ca_3abe
  3D snapshot_portability	(488_334, 0x3f5e_ac53_ad07_5243)	(490_354, 0x504c_043a_1997_c125)
  2D snapshot_portability	(90_574, 0x3019_f8f0_67c5_e643)	(90_960, 0x85b4_446d_161e_81d5)

  The snapshot sizes grow because centered_ball is serialized, so a restored world keeps the flag for recycled steps.

• run-ci-checks.sh passes locally.

• CHANGELOG.md has an Unreleased/Fixed entry.