fix(math_group): replace raw == with MathRubric for math-aware answer comparison

[jfeldstein]

Currently

Both sub-env rubrics in math_group use raw Python string equality to score answers:

return 1.0 if response == answer else 0.0

Semantically equivalent LaTeX forms receive 0 reward instead of 1 — e.g. \frac{1}{2} vs 0.5, \frac{3}{4} vs \dfrac{3}{4}, 1 + x vs x + 1. math-verify was already declared as a dependency in environments/math_group/pyproject.toml but never used.

Change Summary

Two independent TDD cycles, four commits:

Reward function fix (environments/math_group/math_group.py)

• Replace both vf.Rubric(...) calls with vf.MathRubric(parser=parser), consistent with gsm8k, math_python, and doublecheck environments
• Format-reward weights unchanged: gsm8k 0.0 (metric), math 0.2 (weighted)

Teardown propagation fix (verifiers/envs/env_group.py)

• Add EnvGroupRubric.teardown() override to propagate teardown to child rubrics
• MathRubric spawns a ProcessPoolExecutor; previously those workers leaked when the EnvGroup was torn down (vf.Rubric held no resources so this was harmless before this PR)
• Chain: EnvGroupRubric → RubricGroup.teardown() → MathRubric.teardown() → executor shutdown

Acceptance Criteria

• pytest tests/test_math_group.py — 13 tests pass
• pytest tests/test_math_rubric.py tests/test_rubric.py tests/test_rubric_group.py tests/test_env_group.py — existing tests still pass
• \frac{1}{2} vs 0.5, \frac{3}{4} vs \dfrac{3}{4}, 1 + x vs x + 1 all score 1.0
• Wrong answers still score 0.0
• MathRubric executors are shut down when EnvGroup is torn down

Notes

• No new dependencies — math-verify was already in the environment's pyproject.toml
• Metric key changes from math_answer_reward_func / gsm8k_answer_reward_func → correct_answer, consistent with every other math environment in the repo. Both old names are confirmed unused anywhere else in the codebase (ripgrep: 0 matches).

---

Note

Medium Risk
Changes reward scoring semantics for math_group from string equality to math-aware equivalence, which can affect training/eval outcomes. Also adds teardown propagation in EnvGroupRubric, touching lifecycle management and async cleanup paths.

Overview
math_group now uses vf.MathRubric for both gsm8k and math sub-environments, replacing raw string-equality answer checks so symbolically equivalent LaTeX/decimal forms score as correct; the math env still applies a 0.2 weighted format bonus while gsm8k tracks format as a metric.

Adds a EnvGroupRubric.teardown() override to propagate teardown to child environment rubrics (preventing MathRubric executor leaks), and introduces tests/test_math_group.py covering equivalence scoring, format bonus contribution, wrong-answer regression, and teardown propagation.

^{Reviewed by Cursor Bugbot for commit d1bef95. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 679975e. Configure here.}