Fix MJX NaN Issue in Tendon Moment Calculations

[Adityakk9031]

Fixes #2932

Issue
MJX produces NaNs on the very first mjx.step() for models with spatial tendons, while native MuJoCo remains stable. The issue occurs when tendon wrap points are very close or coincident at initialization, causing division by near-zero values.

Root Cause
In
mjx/mujoco/mjx/_src/smooth.py
, the
_length_moment
function computes tendon segment lengths and moments. When the distance between consecutive tendon points is extremely small:

math.safe_div(dif, length) can still produce NaN values
These NaNs propagate through the tendon Jacobian (ten_J)
The constraint solver then propagates NaNs to qacc, and subsequently to
qpos
/
qvel
Solution
Applied a multi-layered fix to
_length_moment
function (lines 910-938):

Denominator Clamping: Replace math.safe_div(dif, length) with dif / jp.maximum(length, mujoco.mjMINVAL) to ensure we never divide by values smaller than the minimum threshold

Degenerate Case Handling: Explicitly track when length < mujoco.mjMINVAL and set moment to zero for these cases

Enhanced Condition: Use jp.logical_or(is_degenerate, body0 == body1) to zero out moments for both degenerate segments and same-body segments

NaN Sanitization: Add conservative guard jp.where(jp.isnan(moment), jp.zeros(m.nv), moment) to catch any remaining numerical issues