Fix MJScene body position pinned at reference pose on mass change

[juan-g-bonilla]

• Tickets addressed: closes MJScene reports body position pinned at the reference pose when a body's mass changes #1476
• Review: By commit
• Merge strategy: Merge (no squash)

Description

MJScene::equationsOfMotion calls mj_setConst whenever a body's mass changes to refresh
MuJoCo's cached "constant" quantities. mj_setConst overwrites data->qpos with the model
reference pose and leaves it there, so every position-dependent quantity computed afterwards in
the same call — forward kinematics, position-dependent forces, and the site state-output
messages — is evaluated at the reference pose instead of the current configuration. Since the
branch fires on every step of a burn, the reported body position stays pinned at the initial
pose for the whole maneuver, even though velocity and mass integrate correctly.

The fix restores the integrator state via updateMujocoArraysFromStates() immediately after
mj_setConst (which also re-flags the kinematics stale so they are recomputed), and clears the
model-const-stale flag now that the constants match the current mass properties.

Commits are organized so the fix is isolated from a behaviour-preserving performance
improvements in the same code path:

1. [#1476] the mj_setConst state-restore fix plus a regression test (MJScene.cpp,
   test_mass_update.py).
2. Cache the mjModel/mjData pointers for the duration of equationsOfMotion instead of
   re-fetching them through the recompile-checking accessor on every access.
3. Release-notes snippet.

Verification

• Extended test_mass_update.py::test_continuouslyChangingMass to also assert the reported
  position against the closed-form trajectory. The existing test only checked velocity, which
  integrates correctly even with the bug present, so it could never have caught this. The new
  assertion requires scene.extraEoMCall = True — with the default extraEoMCall = False,
  postIntegration rewrites the site message after the step and masks the mid-step mj_setConst
  clobber, which is why no existing test surfaced it.
• Confirmed the test fails on the unfixed code (reported x pinned at 0.0 vs analytic 9.66) and
  passes with the fix applied.
• All 15 mujocoDynamics general-module unit tests pass with this branch applied.

Documentation

• Added a release-notes snippet (mjscene-mass-change-and-perf.rst).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8b42c3fa59

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[schaubh]

Good to go.

Maybe update the PR description to reflect the final solution? The description still says “two behaviour-preserving performance improvements,” but only the pointer-caching improvement remains. Change “two … improvements” to “a … improvement.”