[TATP-233]

Summary

• Replaced the hand-rolled inverse-rotation formula in mdp.upright with the shared quat_apply_inverse primitive.
• Kept the quaternion as the read-plan resident tensor and materialized the immutable gravity table directly on env.device.
• Extended the reward regression fixture with a mixed-axis quaternion and non-axis-aligned gravity so the second cross-product term's sign is covered.

For the usual flat-ground gravity vector (0, 0, -1), the old formula's horizontal squared norm happens to match the standard result; the projected vector itself remains wrong. Non-axis-aligned gravity exposes the numerical difference. No in-tree task configuration currently enables mdp.upright, so there is no in-tree task requiring a retraining regression.

Linked Work

• Issue: [Bug?] rewards.py upright 奖励的内联"逆旋转"公式与 quat_apply_inverse 符号不一致 #2585
• Parent roadmap (when applicable): [Roadmap] 统一 math/rotation 原语到 unisim，消除 unilab/unisim 重复实现 #2583
• Roadmap declared base (when applicable): dev/issue-2583-unify-math-rotation (roadmap now merged into main)
• Milestone: none
• Base branch: main

Validation

• make test-all passed on the final local head before this PR was created or updated
• Additional task-specific validation listed below

Commands actually run:

UV_NO_SYNC=1 make test-all

Results:

• Formatting and lint checks passed.
• Mypy passed for 212 source files.
• Pyright reported 0 errors (one pre-existing warning).
• Non-slow coverage suite: 2963 passed, 35 skipped, 375 deselected, 1 xfailed.
• Benchmark smoke: module-mode 35/35 and script-mode 36/36 passed.

Remote CI route:

• Base main: current-head CI link/status: pending at PR creation.

Impact

• Backend impact: none
• Platform impact: both
• Training effect expected: no change for in-tree configured tasks; corrects projected-gravity semantics for any external consumer supplying non-axis-aligned gravity

Artifacts

• W&B: none
• benchmark result: none
• video / screenshot: none
• ONNX / checkpoint: none

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed
• Linked the driving issue
• Noted any follow-up work explicitly