fix(runtime): honor host-bridge manager Torch devices

[TATP-233]

Summary

• Derive ManagerBasedRlEnv/TorchEnv placement from the backend's declared HOST_BRIDGE Torch-device capability. A capable host bridge now uses the current GPU (including ROCm PyTorch's cuda namespace); CPU-only capabilities stay on CPU even when a GPU is visible.
• Keep locomotion gait terms, selected packed resets, policy I/O, and related tests on the negotiated Manager device. Packed tensor-reset event commits now honor the host bridge's declared device contract instead of forcing selected rows through CPU first.
• Update bilingual installation/support documentation and ADR-0012/ADR-0006. The support-matrix generator and its slow test now express that CPU-authoritative host bridges support ROCm Torch buffers without making a CUDA-physics claim.
• Correct the optional Drake startup-boundary assertion to accept the current DrakeEnvPool batch extension wording emitted by this UniSim profile.

CPU-authoritative physics is unchanged. This affects MuJoCo/Motrix/Drake/SuperDex host-bridge execution on Linux CUDA/ROCm and CPU; device-resident backends retain their existing CUDA-only lifecycle.

Linked Work

• Issue: fix(runtime): honor host-bridge Torch devices on ROCm #2083
• Parent roadmap (when applicable): roadmap: tensor-only Manager-Based runtime and three-backend scope reduction #1811
• Roadmap declared base (when applicable): 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_NO_SYNC=1 uv sync --all-extras --dev
UV_NO_SYNC=1 make check
UV_NO_SYNC=1 uv run pytest -q \
  tests/envs/test_manager_based_rl_env.py \
  tests/envs/locomotion/test_manager_gait_terms.py \
  tests/envs/locomotion/go2/test_manager_based_cfg.py \
  tests/base/test_entity_scene_consumer.py \
  tests/managers/test_isaac_lab_migration_fixture.py \
  tests/managers/test_mjlab_migration_fixture.py \
  tests/managers/test_stage_curriculum_demo.py
UV_NO_SYNC=1 uv run pytest -q \
  tests/envs/test_go2_superdex.py \
  tests/envs/locomotion/go2/test_manager_based_cfg.py::test_go2_flat_drake_factory_is_registered_before_optional_runtime_load \
  tests/scripts/test_check_docs.py::test_documentation_files_match_current_repo_contracts
UV_NO_SYNC=1 uv run scripts/generate_support_matrix.py --write
UV_NO_SYNC=1 uv run pytest -q -m slow tests/scripts/test_support_matrix.py
UV_NO_SYNC=1 make test-all

Final local result:

• make check: passed (ruff format, ruff check, mypy, pyright, focused test lint).
• make test-all: passed (1854 passed, 17 skipped, 489 deselected, benchmark smoke imports 34/34 module-mode and 35/35 script-mode).
• Focused GPU coverage exercised CUDA host-bridge placement and SuperDex CUDA packed resets on the local GPU. CPU-only capability fallback is also covered.

Remote CI route:

• Base main: current-head remote CI requested below; status will be recorded after it reports.

Impact

• Backend impact: mujoco / motrix / drake / superdex (HOST_BRIDGE only; device-resident paths unchanged)
• Platform impact: Linux (CPU, CUDA, ROCm) / macOS CPU-authoritative host bridge; no new CUDA-only physics claim
• Training effect expected: yes — GPU host-bridge Manager/policy I/O stays on the negotiated GPU instead of being forced to CPU; dimensions and policy contracts are unchanged

Artifacts

• W&B: N/A
• benchmark result: N/A (benchmark smoke included in make test-all)
• video / screenshot: N/A
• ONNX / checkpoint: N/A

Checklist

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

No unresolved follow-up is known. The remote CUDA/ROCm-specific hardware coverage remains dependent on CI runners.

[TATP-233]

CI link for the current head: https://github.com/Motphys/UniLab/actions/runs/37776382218

[TATP-233]

Final-head CI for fa89b76e: https://github.com/Motphys/UniLab/actions/runs/37779801809

[TATP-233]

Correction: final-head CI run is https://github.com/Motphys/UniLab/actions/runs/37776693104 for fa89b76e.

[TATP-233]

Current-head CI for c3e0d593: https://github.com/Motphys/UniLab/actions/runs/37778027767 (previous CPU-only-Torch CI failure was fixed).

[TATP-233]

Current-head CI for ae7396b3: https://github.com/Motphys/UniLab/actions/runs/37779358883. CPU-only Torch builds now skip the fake-GPU negative branch.

[TATP-233]

Current-head CI passed: https://github.com/Motphys/UniLab/actions/runs/37779358883. Current-head Docs passed: https://github.com/Motphys/UniLab/actions/runs/37779358609.