[TATP-233]

Summary

• Encode the angular channels of community 13-D world-frame root state before fused tensor reset submission.
• write_root_state_tensor() and write_motion_state_tensor() now share a device-resident encoder: linear velocity stays world-frame, while angular velocity is rotated with quat_apply_inverse(root_quat, w_world) into generalized free-root/body-frame qvel.
• Keep the existing fused submission structure and backend commit boundary. The encoder uses one reusable full-width scratch tensor, so it does not return to the host reset path or add stable allocations per reset.
• This corrects immediate reset state and downstream MotionCommand motion resets for rotated roots and nonzero world angular velocity.

Linked Work

• Issue: Bug: fused tensor root reset 未把世界系角速度编码为 body-frame free-root qvel #2512
• Parent roadmap: N/A
• Roadmap declared base: N/A
• Milestone: N/A
• 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 run pytest tests/base/test_reset_state.py -q
# 39 passed

uv run pytest tests/base/test_backend_conformance.py -q -k 'root_qvel_body_angular or fused_tensor_root_reset'
# 2 passed, 2 skipped (Motrix unavailable locally; MuJoCo authoritative readback passed)

uv run pytest tests/tasks/test_motion_term_parity.py::test_motion_reset_values_kernel_matches_eager_and_reuses_scratch -q
# 1 passed

uv run ruff format --check src/unilab/base/reset_state.py tests/base/test_reset_state.py tests/base/test_backend_conformance.py
uv run ruff check src/unilab/base/reset_state.py tests/base/test_reset_state.py tests/base/test_backend_conformance.py
# passed

uv run mypy src/unilab
uv run pyright
# 0 errors

make test-all
# formatting/types/test checks passed
# 2781 passed, 61 skipped, 373 deselected, 24 warnings
# benchmark smoke: module mode 34/34, script mode 35/35

Remote CI route:

• Base main: current-head CI is requested with this PR; status will be checked after GitHub schedules it.

Impact

• Backend impact: mujoco / motrix / both
• Platform impact: both
• Training effect expected: yes — motion and tensor root resets with rotated roots and nonzero world angular velocity now initialize the intended physical angular velocity.

Artifacts

• W&B: N/A
• benchmark result: N/A (no benchmark regressions introduced; make test-all benchmark smoke passed)
• video / screenshot: N/A
• ONNX / checkpoint: N/A

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed (no public API semantic change; this restores the documented world-frame input contract)
• Linked the driving issue
• Noted any follow-up work explicitly

Follow-up: run the Motrix parameterized conformance regression in an environment with motrixsim; it was skipped locally because that optional runtime is not installed.