fix(env): commit MuJoCo model-field DR through host bridge

[TATP-233]

Summary

• Materialize the backend before startup model-field DR. Startup body_mass / body_com still sample once, commit full-width through the public SimBackend.set_state(..., randomization=...) boundary, and run before read-plan binding, the first reset, or a step. MuJoCo no longer attempts to commit startup DR before its public selected-reset/model pool exists.
• Carry host-staged model-field DR through the packed HOST_BRIDGE tensor-event reset boundary. A tensor command carrier and body_mass, body_com (body_ipos), geom_friction, or PD gains now compose into one HostBridgeTransferPlan.apply_reset(..., randomization=...) call.
• Fail closed when a backend does not declare TensorLifecycleCapabilities.reset_randomization; mapped scene patches remain unsupported on this boundary.
• Add fake-boundary lifecycle/payload regressions plus a real MuJoCo FlashSAC G1 tensor-command regression for both startup and reset DR.
• Update the English and Chinese Manager-Based architecture and DR contract documents for the revised lifecycle ordering and packed HOST_BRIDGE reset DR contract.

User-facing impact: downstream manager-based MuJoCo tasks using the tensor command carrier / HOST_BRIDGE path no longer need to disable the issue's startup/reset model-field DR terms.

Linked Work

• Issue: Bug: MuJoCo backend cannot commit startup or reset model-field DR with tensor command/HOST_BRIDGE path #2122
• Parent roadmap (when applicable): none
• Roadmap declared base (when applicable): none
• 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:

make check
# Passed: ruff format/check, mypy, pyright, and test F401/F821/F811/F841 lint.

make test
# 1898 passed, 42 skipped, 497 deselected, 11 warnings.

make test-all
# Passed.
# Non-slow coverage suite: 1898 passed, 42 skipped, 497 deselected, 11 warnings.
# Benchmark smoke: module-mode 34/34; script-mode 35/35.

uv run pytest tests/envs/test_manager_based_rl_env.py tests/base/test_reset_state.py tests/envs/mdp/test_events.py -q
# 200 passed, 1 warning.

uv run pytest tests/envs/test_env_configs.py::test_flashsac_g1_motion_mujoco_tensor_command_model_field_dr_roll_out \
  -q -m slow --override-ini 'addopts='
# 2 passed, 1 existing non-writable NumPy array warning.

Remote CI route:

• Base main: current-head CI link and status recorded below after the run starts.

Impact

• Backend impact: mujoco (public HOST_BRIDGE lifecycle only; no backend-private access)
• Platform impact: Linux
• Training effect expected: yes — enables startup/reset model-field DR on the MuJoCo tensor command / HOST_BRIDGE path without changing the default task-owner DR configuration.

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

Follow-up: if a future backend needs to distinguish scalar set_state DR support from packed host-bridge apply_reset DR support, add a dedicated public capability rather than probing backend internals.

[TATP-233]

Remote CI on current head ca5a7eabee7df48f79c5cb03a7fae435f4957ea6 completed successfully: CI run 38042119193 — ruff-lint, ruff-format, mypy, pyright, benchmark-smoke, and test (ubuntu-slim) all passed. Docs build also passed: Build Sphinx.