[adenzler-nvidia]

The noslip solver restores a contact's old friction forces when an update raises the cost by more than 1e-10, which is MuJoCo's double-precision threshold. In float32 that is below the rounding of the new forces: on an active cone row the gradient is not zero, so one ulp of force moves the cost by about eps * |force * gradient|, a few 1e-7 in the test below. Near convergence the guard therefore rejects valid updates depending on rounding, and because a rejected update leaves the state unchanged, every later sweep rejects it again and the contact stalls short of MuJoCo's solution. This showed up as a CPU-only failure of test_noslip43 in #1879, which only changes the main solver's rounding.

This adds that rounding bound to the threshold in the pyramidal and elliptic guards, with the same 8 * eps factor the line search uses. Evaluating the cost change in float64 does not help, because the noise is in the rounded forces and not in the arithmetic. A failed QCQP returns zero force, for which the bound is zero, so those updates are still restored as before.

The new test sweeps the spin of a sphere on a plane (elliptic, condim 6) over 61 worlds; on main some of these worlds miss MuJoCo's qacc by up to 1.9e-4 relative. In the pyramidal guard the rejected updates I found were only 1-2 ulps, so that change is for consistency and has no test of its own.

humanoid at 8192 worlds on an RTX PRO 6000 Blackwell, main → this PR in steps/s: 2.009M → 2.010M with noslip_iterations=3, 1.919M → 1.918M with elliptic cones as well, 9.128M → 9.125M with noslip off.

[thowell]

LGTM! Thanks @adenzler-nvidia -- nice catch and clean fix scaling the noslip cost-change guard to float32 force-rounding noise while preserving zero-force rollback on failed QCQPs.