[envpool] Upgrade tests to Python 3.12 and MuJoCo 3.x

[Trinkle23897]

Summary

• Problem: Bazel test and dependency plumbing was still anchored to Python 3.10 and MuJoCo 2.2.x, which blocked Python 3.12 coverage and newer Gymnasium MuJoCo env ids.
• Scope: Upgrade the Bazel Python toolchain and pip dependency plumbing for 3.12, expand declared support to 3.10-3.13, migrate MuJoCo gym coverage to v5/3.x, and keep existing tests running without introducing new skip/xfail paths.
• Outcome: The runtime/toolchain updates and test fixes are in this PR; the GitHub workflow YAML edits are parked under manual_restore/workflows/ so they can be moved back from a machine with workflow scope.

This keeps the Python 3.12 bring-up and the MuJoCo 3.x migration in one coherent update, while preserving the existing test surface instead of hiding failures behind new skips.

Technical Details

• Approach: Bump rules_python to a WORKSPACE-compatible 0.27 release, switch Bazel pip setup from pip_install to pip_parse with dev/release lockfiles, vendor MuJoCo 3.3.4 plus Gymnasium v5 XML overlays, and update the MuJoCo/VizDoom tests so they still exercise the same paths under the newer dependencies.
• Code pointers:
  • WORKSPACE, envpool/workspace0.bzl, envpool/workspace1.bzl, envpool/pip.bzl: Python 3.12 toolchain registration, rules_python upgrade, and lockfile-based pip setup.
  • envpool/mujoco/gym/registration.py, envpool/mujoco/gym/*.h, third_party/mujoco_gym_xml_patches/*: MuJoCo 3.x backend updates and Gymnasium v5 env/XML alignment.
  • envpool/mujoco/*align*_test.py, envpool/vizdoom/vizdoom_pretrain_test.py: test updates; no new skip/xfail markers were introduced, and test_d3 keeps its quality gate while running in a subprocess.
  • manual_restore/workflows/*: workflow YAML parked here after PushPatrol rejected direct .github/workflows/* updates without workflow scope.
• Notes: The repo already had separate dev/release requirements; the new requirements-dev-lock.txt and requirements-release-lock.txt files preserve that split under newer rules_python.

Test Plan

Automated

• git diff --unified=0 origin/main -- . ':!third_party/pip_requirements/requirements-*-lock.txt' | rg -n "skip|xfail|pytest\.mark\.skip|self\.skipTest|@unittest\.skip|absltest\.skip": no new skip/xfail markers
• git diff --check: pass
• brix ssh dev -C -- 'cd /root/code/envpool && bazel test //envpool/mujoco:mujoco_gym_align_test //envpool/mujoco:mujoco_gym_deterministic_test //envpool/vizdoom:vizdoom_pretrain_test --test_output=errors --cache_test_results=no --config=test': pass
• brix ssh dev -C -- 'cd /root/code/envpool && bazel test //envpool/vizdoom:vizdoom_pretrain_test --test_output=errors --cache_test_results=no --config=test': pass after restoring the D3 quality gate
• brix ssh dev -C -- 'cd /root/code/envpool && make bazel-test': pass (30/30)

Suggested Manual

• Move manual_restore/workflows/lint.yml -> .github/workflows/lint.yml and manual_restore/workflows/release.yml -> .github/workflows/release.yml from a machine with workflow scope, then push the branch again
• make bazel-build: recheck the build path separately, since this series focused on test/runtime compatibility and not the wheel build path

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 99b6e27b9b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Thread xml_file through the other v5 MuJoCo envs

Gymnasium v5 restored the xml_file kwarg across the MuJoCo suite, but here only Pusher and Walker2d get it. HopperEnv, AntEnv, HumanoidEnv, HumanoidStandupEnv, Inverted*, Reacher, HalfCheetah, and Swimmer still hard-code their asset path, so calls like make_gymnasium('Hopper-v5', xml_file='custom.xml') will fail during config generation instead of loading the custom model. That blocks one of the main v5 features for most of the newly added env IDs.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Emit the Gymnasium v5 info fields for Hopper/Walker2d

This upgrades Hopper/Walker2d to the v5 reward semantics, but WriteState() still only exports x_position and x_velocity. Gymnasium v5's info for these envs also includes the reward breakdown (reward_forward, reward_ctrl, reward_survive) plus z_distance_from_origin, so downstream logging and wrappers that expect the standard v5 info dict will silently lose those fields when switching to EnvPool's -v5 registrations. envpool/mujoco/gym/walker2d.h has the same WriteState(reward, xv, x_after) pattern.

Useful? React with 👍 / 👎.