feat(macos): enable mujoco/mjbatch tensor-runtime off-policy training on macOS

[TATP-233]

Summary

Closes the UniLab portion of roadmap #2055: run the completed
npenv→torchenv runtime on macOS (Apple Silicon) for the MuJoCo/mjbatch
HOST_BRIDGE backend, with SAC + g1_walk_flat and FlashSAC +
g1_motion_tracking training on CPU physics + MPS learner.

• chore(workspace): realign tensor_runtime_workspace.json sibling pins
  with the develop/tensor-runtime branch tips. UniLab HEAD already consumes
  the packed-reset public-width contract
  (SimBackend.get_public_state_widths / PublicStateWidths) that only
  exists at the tips; the stale pins made
  ResetStateTransaction.can_commit_packed() fail closed on mujoco, leaving
  the host-bridge tensor reset dispatch (added in 24e57ffa) without a viable
  packed commit and breaking flashsac + g1_motion_tracking --sim mujoco at
  the first autoreset. Sibling integration branches
  dev/issue-2055-macos-tensor-runtime are based on those exact tips; no
  sibling code change was required.
• test(platform): the fake-runtime / mock-worker CUDA-only backend suites
  bypass the factory platform gate (dedicated gate coverage stays in
  tests/base/test_cuda_backend_platform_preflight.py), restoring 58 failed
  • 18 errored adapter tests on all CUDA-less hosts; plus targeted skips
    (CUDA observation naming, mjwarp-extra soak command, Linux-/proc soak
    resource monitor) and removal of the never-committed
    outputs/velb_compare/micro_velb.py benchmark-scope entry.
• test(g1): new slow-lane regression case — selected-row reset through the
  packed host-bridge tensor motion command commit on mujoco; newton G1 owner
  cases skip off-CUDA.

Platform / behavior impact

macOS + MuJoCo (mjbatch) only. No change to runtime code, contracts, CUDA
behavior, Linux behavior, or training semantics: the only non-test change is
the workspace manifest realignment. No npenv-runtime / np-rng revert. No
main or develop/tensor-runtime mutation. No PyPI publication.

Validation (exact commands on this head, macOS arm64, torch 2.14.0 MPS)

• make check — pass (ruff, mypy 144 files, pyright 0 errors).
• make test (uv run pytest -m "not slow" -q) — 1801 passed, 59 skipped.
• make test-all — pass (check + coverage lane + benchmark smoke 35/35 and 36/36).
• Focused slow lane:
  • uv run pytest tests/envs/test_env_configs.py -m slow -k mujoco — 5 passed (incl. new selected-reset case).
  • uv run pytest tests/envs/locomotion/g1/test_g1_owner_contract.py -m slow — 17 passed, 4 skipped.
  • uv run pytest tests/tasks/test_flashsac_owner_contract.py — 10 passed.
• Training smokes (algo.max_iterations=100, training.no_play=true):
  • uv run train --algo sac --task g1_walk_flat --sim mujoco — 100/100, normal_completion, collector exit 0.
  • uv run train --algo flashsac --task g1_motion_tracking --sim mujoco — 100/100, normal_completion, collector exit 0.
• Acceptance full run:
  • uv run train --algo sac --task g1_walk_flat --sim mujoco training.play_render_mode=record
    (owner-default algo.max_iterations=5000) — completed 5000/5000,
    10,260,480 env steps in 11.8 min (~14.3k steps/s), final mean reward
    278.98 / best 303.60 / episode length 946.85, normal_completion,
    collector exit 0, no residual processes; ONNX export verified
    (max_diff 7.8e-07); 800-frame playback video rendered.

Raw evidence: issue #2055 comments.

Governance

Declared base: develop/tensor-runtime. Related: #1706 (NpEnv removal),
#2052 (Newton canonical tensor runtime), 5-contributing_workflow.md
tensor-runtime sibling dependency transition (profile stays local-editable;
publication remains out of scope). No new public contract, no ADR required:
the packed-reset width contract is declared by the sibling branch tips this
manifest now pins.