fix(training): stabilize APPO and MJWarp tests

[TATP-233]

Summary

• Fixed APPO HOST_BRIDGE device routing so only an explicit CUDA collector request or explicit runner/config device selects CUDA Manager Torch carriers. Device auto-detection no longer silently puts the CPU-authoritative MuJoCo Manager runtime on CUDA.
• Made the APPO G1 motion-tracking config-load smoke deterministic with an explicit CPU learner device, independent of whether the CI/host machine exposes CUDA.
• Stabilized the MJWarp low-friction DR regression by moving slide friction from the numerically pathological 0.01 boundary to 0.1; the test still verifies that materially lower friction changes kick response while same-condition worlds remain pairwise consistent.

Linked Work

• Issue: fixes the reported test_appo_task_configs_load[g1_motion_tracking/mujoco] startup failure and test_low_friction_world_changes_kick_response flake. Upstream IPC fix: fix(ipc): bound APPO rollout tensor views to payload size on macOS unilabsim/unilab_rl#91
• 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:

# final rebased head
uv run --no-sync pytest \
  tests/base/test_mjwarp_dr.py::test_low_friction_world_changes_kick_response \
  "tests/scripts/test_train_script_configs.py::test_appo_task_configs_load[g1_motion_tracking/mujoco]" \
  tests/scripts/test_train_scripts.py::test_appo_host_bridge_carriers_require_explicit_cuda \
  -q --no-header -m slow
# 3 passed

make test-all
# ruff format: 397 files unchanged
# ruff check: all checks passed
# mypy: no issues in 133 source files
# pyright: 0 errors, 0 warnings, 0 informations
# focused tests lint check: passed
# non-slow coverage suite: 1885 passed, 21 skipped
# benchmark smoke: module-mode 34/34; script-mode 35/35

Flake evidence and focused stability:

# Original friction value: NaN reproduced once in 10 runs (world 3 became non-finite).
# Updated value: 20 consecutive focused runs passed, followed by the final reruns above.

Remote CI route:

• Base main: current-head CI will be recorded on this PR and must pass before merge.

Impact

• Backend impact: mujoco (APPO smoke), mjwarp (test-only friction value)
• Platform impact: Linux validated; APPO fix also prevents CUDA-dependent behavior on hosts without CUDA
• Training effect expected: yes — restores deterministic APPO G1 startup for the CPU-carrier owner without changing explicit CUDA collector behavior

Artifacts

• W&B: N/A
• benchmark result: N/A
• video / screenshot: N/A
• ONNX / checkpoint: N/A

Checklist

• Added or updated tests where needed
• Updated docs if behavior or workflow changed (internal routing/test behavior only; no user-facing workflow change)
• Linked the driving issue
• Noted any follow-up work explicitly

[TATP-233]

Current-head remote CI: https://github.com/Motphys/UniLab/actions/runs/37807084690 — lint, type, docs, and benchmark jobs passed. The job was still reported as in progress when the run rollup completed; I’m tracking it and will update this PR before merge.