feat(rl): add WarpSAC G1 owners for unilab-rl 1.4.0

[TATP-233]

Summary

• Pinned standard and ROCm dependencies/locks to unilab-rl==1.4.0; no release or PyPI publication is included.
• Added the WarpSAC entrypoint/config tree, structured config, runner dispatch, CLI/completion/docs/support-matrix integration, and 1.4.0 field cleanup.
• Added and tuned WarpSAC owners for g1_walk_flat on MuJoCo/mjwarp and g1_motion_tracking on MuJoCo/mjwarp.
• Fixed explicit post-training playback run selection so ONNX/video are generated from the final checkpoint rather than a stale algo.load_run selector.

Linked Work

• Closes Add validated WarpSAC G1 owners for unilab-rl 1.4.0 #1644
• Governing ADR: docs/sphinx/source/adr/ADR-0003-task-owner-and-config-compose-contract.md
• Base branch: main
• Milestone: none

Tuning and full-run evidence

Tuning decisions:

• Walk: decay_step=2048, replay_min_weight=0.10, actor/critic parameter normalization disabled.
• Motion: uniform replay (decay_step=0), replay_min_weight=0.05, actor/critic parameter normalization enabled.

All four final WarpSAC runs completed with final checkpoints, verified ONNX, and video:

Task	Backend	Env steps	Final reward	Best reward	Episode length	Training wall
Walk	MuJoCo	20,680,704	296.136	298.545	993.12	167.77 s
Walk	mjwarp	20,680,704	299.591	300.279	998.94	158.62 s
Motion	MuJoCo	51,400,704	26.847	32.470	290.49	486.92 s
Motion	mjwarp	51,400,704	28.915	32.868	278.36	582.10 s

Existing FastSAC/FlashSAC logs were analyzed without rerunning them. Full tables and comparability caveats are in logs/warpsac_comparison/REPORT.md; curves use cumulative environment steps and reward/mean_ep100:

• logs/warpsac_comparison/g1_walk_mujoco_warpsac_comparison.png
• logs/warpsac_comparison/g1_walk_mjwarp_warpsac_comparison.png
• logs/warpsac_comparison/g1_motion_mujoco_warpsac_comparison.png
• logs/warpsac_comparison/g1_motion_mjwarp_warpsac_comparison.png

Walk caveat: FastSAC uses 2,048 envs, 10.260M steps, and a different reward-owner profile. Its curve is only a common-prefix reference; FlashSAC/WarpSAC are directly comparable at 20.681M steps. Motion runs all use 2,048 envs and approximately 51.2–51.4M steps.

Final artifacts:

• Walk MuJoCo: logs/warp_sac/G1WalkFlat/final_fba92b8f_mujoco/{policy.onnx,play_video.mp4}
• Walk mjwarp: logs/warp_sac/G1WalkFlat/final_fba92b8f_mjwarp/{policy.onnx,play_video.mp4}
• Motion MuJoCo: logs/warp_sac/G1MotionTrackingSAC/final_ddb2908b_mujoco/{policy.onnx,play_video.mp4}
• Motion mjwarp: logs/warp_sac/G1MotionTrackingSAC/final_ddb2908b_mjwarp/{policy.onnx,play_video.mp4}

ONNX checker results: walk [1,98] -> [1,29]; motion [1,160] -> [1,29]. ffprobe verified every video as H.264, 1280x720, 800 frames at 50 fps (16 s).

Validation

• make test-all passed on the final local head before this PR was created
• Additional task-specific validation listed below

Commands and outcomes:

uv run pytest --override-ini='addopts=--tb=short' \
  tests/visualization/test_interactive_playback.py::test_sac_playback_session_runs_sim2sim_preflight \
  tests/config/test_locomotion_params.py::test_offpolicy_warpsac_g1_task_overrides
# 2 passed

make check
# ruff, mypy, pyright, and focused test lint checks passed

make test
# 1548 passed, 26 skipped, 604 deselected, 12 warnings

make test-all
# check passed; 1548 passed, 26 skipped, 604 deselected;
# benchmark module smoke 35/35 and script smoke 36/36 passed

uv run scripts/audit_sim2sim_contracts.py --trees warpsac
# 0 transferable, 0 blocked, 0 guard-blind-spot fields

uv lock --check
# standard lock resolved and valid

# ROCm files were temporarily swapped into pyproject.toml/uv.lock, checked, then restored:
uv lock --check
# ROCm lock resolved and valid

uv tree --package unilab-rl
# unilab-rl v1.4.0

The unrelated ignored scripts/benchmark/outputs/ tree was temporarily moved outside the repo for make test/make test-all and restored afterward; its docs-check contents are unrelated local benchmark output.

Experimental validation used local, pc823, and pc825 in parallel. Remote pc825 was updated to the final playback-fix head before regenerating its final motion ONNX/video from model_25000.pt.

Remote CI route:

• Base main: CI https://github.com/Motphys/UniLab/actions/runs/36160010924 — success at commit 91407a59725fef001293b169318d3a951863d8fd (all six jobs passed).
• Docs: https://github.com/Motphys/UniLab/actions/runs/36160010928 — success at commit 91407a59725fef001293b169318d3a951863d8fd.

Impact

• Backend impact: MuJoCo and mjwarp training/playback paths; other backends unchanged.
• Platform impact: Linux (training validation used CUDA); configuration/docs are cross-platform.
• Training effect expected: yes, adds WarpSAC and tuned G1 owners.

Artifacts

• W&B: none (TensorBoard only)
• Benchmark result: logs/warpsac_comparison/summary_metrics.json
• Video / screenshot: four final play_video.mp4 files listed above
• ONNX / checkpoint: four final policy.onnx files and run checkpoints listed above

Checklist

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