Cherry-pick Antoine PRs to release/3.0.0-beta2

[AntoineRichard]

Summary

Cherry-picked these merged PRs from develop to release/3.0.0-beta2:

• Fix transformation matrix docstrings on develop #5895 Fix transformation matrix docstrings on develop
• Fix cloner lazy exports #5920 Fix cloner lazy exports
• Enable Spot Newton MJWarp preset #5940 Enable Spot Newton MJWarp preset
• Pin ovphysx runtime #5959 Pin ovphysx runtime
• Remove obsolete warp-to-torch migration helper #5924 Remove obsolete warp-to-torch migration helper
• Fix pre-trained policy action grad leak #5923 Fix pre-trained policy action grad leak

Release-Branch Notes

• For Pin ovphysx runtime #5959, kept the release branch setup.py packaging path and applied the ovphysx==0.4.13 pin there.
• Added the missing isaaclab_tasks changelog fragment for the Enable Spot Newton MJWarp preset #5940 Spot preset change.

Verification

• ./isaaclab.sh -f
• ./isaaclab.sh -p tools/changelog/cli.py check release/3.0.0-beta2
• ./isaaclab.sh -p -m pytest source/isaaclab/test/test_scripts_torcharray_patterns.py
• ./isaaclab.sh -p -c "from pathlib import Path; import ast; ast.parse(Path('source/isaaclab/isaaclab/cloner/__init__.pyi').read_text()); print('cloner stub syntax ok')"

[isaaclab-review-bot]

🤖 Isaac Lab Review Bot

Summary

Cherry-pick of 5 PRs from develop to release/3.0.0-beta2:

• #5940 – Enable Spot Newton MJWarp preset (Newton collision pipeline, increased capacity/margin)
• #5959 – Pin ovphysx==0.4.13 runtime dependency
• #5924 – Remove obsolete wrap_warp_to_torch.py migration helper
• #5956 – Deprecate MJWarpSolverCfg.ls_parallel (kept as ignored compat field with deprecation warning)
• #5923 – Fix pre-trained policy action grad leak (benchmark wrapped with torch.inference_mode())

The PR also notes a cherry-pick conflict resolution for #5959: the release branch setup.py packaging path was preserved while applying the ovphysx==0.4.13 pin.

Cherry-Pick Correctness

• ✅ No merge conflict markers (<<<<<<<, =======, >>>>>>>) found in the diff
• ✅ All 33 files changed cleanly — removals of ls_parallel usage are consistent across all task configs, tests, and documentation
• ✅ The ls_parallel field is properly deprecated (kept with __post_init__ warning + ignored in _build_solver) rather than hard-removed, preserving backward compatibility
• ✅ Changelog fragments added for all affected packages (isaaclab, isaaclab_newton, isaaclab_contrib, isaaclab_tasks, isaaclab_tasks_experimental, isaaclab_ovphysx)
• ✅ Documentation updates are internally consistent (migration guide, solver docs, solver comparison all aligned)
• ✅ CI workflow pins match the setup.py pin (ovphysx==0.4.13)
• ✅ OVPhysX test docstrings updated from version-specific ovphysx<=0.3.7 to generic "The OVPhysX runtime" phrasing

Findings

No issues found. The cherry-pick is clean with no conflict artifacts. The conflict resolution for #5959 (keeping release branch packaging paths while applying the version pin) is correctly described and the changes are consistent. All ls_parallel removals from call sites are paired with the deprecation infrastructure in the cfg class, ensuring no breakage for downstream code that still passes the parameter.

Verdict

No issues found — clean cherry-pick ready for merge.

---

Update (cd7d0e9): 4 additional files changed since last review:

1. scripts/benchmarks/benchmark_startup.py — Wraps env.step(actions) in torch.inference_mode() context (part of #5923 fix, consistent with PR intent)
2. source/isaaclab/isaaclab/utils/math.py — Docstring correction: T_{AB} described as "frame B to A" instead of "frame A to B" (matches actual convention used in code — t01 = position of frame 1 w.r.t. frame 0, i.e. transform from 1→0)
3. source/isaaclab/isaaclab/cloner/__init__.pyi — Fixes lazy exports: removes cfg_source_path/path_source_path, adds resolve_clone_plan_source (API rename reflected in stub)
4. source/isaaclab/changelog.d/ — Two new changelog fragments (docstring-only skip + cloner fix entry)

✅ No issues. Changes are consistent with original PR scope (cherry-picks from develop). The math docstring fix corrects a genuine documentation error.

---

Update (99d683f): 1 file added since last review:

1. source/isaaclab_tasks/changelog.d/antoiner-spot-newton-mjwarp-preset.rst — Changelog fragment for the Spot Newton MJWarp preset fix (documents the collision pipeline correction from #5940)

✅ No issues. Trivial changelog addition consistent with existing PR scope.

[greptile-apps]

Greptile Summary

This cherry-pick from develop bundles five PRs for the release/3.0.0-beta2 branch: deprecation of MJWarpSolverCfg.ls_parallel (now a no-op that emits a DeprecationWarning), removal of the wrap_warp_to_torch.py migration tool, pinning ovphysx==0.4.13 across CI and packaging, an updated Spot locomotion Newton preset, and a torch.inference_mode() guard in the startup benchmark.

• ls_parallel deprecation: __post_init__ warns and resets the field to False; mjwarp_manager.py explicitly excludes it from kwargs forwarded to SolverMuJoCo. All task/test configs that set it (including ls_parallel=True in Ant and Humanoid) have the field removed.
• ovphysx pin: setup.py, build.yaml, and daily-compatibility.yml are all updated consistently to ovphysx==0.4.13 to guard against a known-breaking newer release; the strict equality pin is intentional per the PR description.
• Spot env update: njmax raised from 45 → 130, nconmax from 30 → 40, switched to Newton collision pipeline (use_mujoco_contacts=False) with a NewtonCollisionPipelineCfg and NewtonShapeCfg(margin=0.01).

Confidence Score: 4/5

Safe to merge; all changes are targeted deprecations, dependency pins, and config updates with no risk of data loss or crash.

The deprecation path for ls_parallel is implemented consistently across config, manager, tests, and docs. The ovphysx==0.4.13 exact-equality pin is intentional to block a known breaking release and is mirrored across all three affected files. The only minor concern is that the strict == pin will block future patch releases of ovphysx, which may need revisiting as the library evolves.

source/isaaclab_ovphysx/setup.py — the exact-equality pin for ovphysx may need to be relaxed once a known-good patch release is available.

Important Files Changed

Filename	Overview
.github/workflows/build.yaml	Pins ovphysx to ==0.4.13 in extra-pip-packages for rendering correctness test job, consistent with setup.py change.
.github/workflows/daily-compatibility.yml	Pins ovphysx to ==0.4.13 in extra-pip-packages for daily compatibility job, consistent with setup.py change.
source/isaaclab_ovphysx/setup.py	Pins ovphysx extras_require dependency to exactly ==0.4.13 to guard against a breaking newer release; strict equality pin blocks any future patch releases.
source/isaaclab_newton/isaaclab_newton/physics/mjwarp_manager_cfg.py	Deprecates ls_parallel via post_init with a DeprecationWarning; stacklevel=5 is correct given the configclass _combined_function wrapper adds one extra frame.
source/isaaclab_newton/isaaclab_newton/physics/mjwarp_manager.py	Explicitly adds ls_parallel to the ignored set when building kwargs for SolverMuJoCo, ensuring the deprecated field is never forwarded to the underlying solver.
scripts/tools/wrap_warp_to_torch.py	Entire 648-line migration helper deleted; changelog entry correctly documents users should use ProxyArray.torch for tensor interop instead.
scripts/benchmarks/benchmark_startup.py	Wraps env.step(actions) in torch.inference_mode() for the first-step profile, avoiding gradient-tracking overhead in the benchmark measurement.
source/isaaclab_tasks/isaaclab_tasks/manager_based/locomotion/velocity/config/spot/flat_env_cfg.py	Significantly updates Spot Newton preset: njmax 45→130, nconmax 30→40, adds use_mujoco_contacts=False, NewtonCollisionPipelineCfg(max_triangle_pairs=2_500_000), and NewtonShapeCfg(margin=0.01).

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["User creates MJWarpSolverCfg(ls_parallel=True)"] --> B["@configclass __init__"]
    B --> C["_combined __post_init__\n(user + _custom_post_init)"]
    C --> D{"ls_parallel == True?"}
    D -- Yes --> E["warnings.warn(DeprecationWarning)\nstacklevel=5"]
    E --> F["Set ls_parallel = False"]
    D -- No --> G["No-op"]
    F --> H["NewtonMJWarpManager.__init__"]
    G --> H
    H --> I["ignored = {class_type, solver_type, ls_parallel}"]
    I --> J["kwargs = cfg.to_dict() filtered by SolverMuJoCo signature - ignored"]
    J --> K["SolverMuJoCo(model, **kwargs)"]

Loading

_{Reviews (1): Last reviewed commit: "Fix pre-trained policy action grad leak ..." | Re-trigger Greptile}

[greptile-apps]

The exact == pin means any ovphysx version other than 0.4.13 — including future bug-fix releases like 0.4.14 — will be refused by pip. A compatible-release (~=0.4.13) or upper-bounded range (>=0.4.13,<0.5) would block known-breaking minor bumps while still allowing patch releases.

Suggested change

	"ovphysx": ["ovphysx==0.4.13"],
	"ovphysx": ["ovphysx~=0.4.13"],

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!