[burtenshaw]

Summary

Spawned QED verifier workers eagerly imported the client and complete server, including Gradio and datasets, inside the five-second request deadline. Cold startup can exhaust both attempts and turn correct answers or wrong-answer penalties into neutral timeout rewards (seen in CI on #1469 and #1270).

 worker import
-  client + OpenEnv/Gradio + server/datasets/OpenAI
+  math verification dependencies only
 public package exports
+  resolve lazily, preserving class identities

Only the two QED package initializers change. Request deadlines, retries, grading, and existing fixtures are unchanged.

Evidence

• Before: Python 3.11 cold worker import: 13.907s. Injecting an 11s delay into datasets import reproduces test_correct_gives_max_reward: score 0 instead of 7, with two 5s timeouts; workers subsequently finish. The new subprocess regression fails because the worker imports openenv.
  After: worker import: 0.216s. All three originally failing tests pass with that same delay; 149 QED tests pass (5 integration tests deselected). Regression verifies isolated imports, real spawned correct/wrong grading, and public export identity.

• Relevant dependency versions match failing CI: math-verify 0.9.0, SymPy 1.14.0, latex2sympy2-extended 1.11.0, ANTLR 4.13.2. Reproduction is local macOS Python 3.11; CI itself has not been instrumented.

• Changed-file formatting/import sorting and repository ruff check src/ tests/ pass. Global formatting hook reports existing issues (including grid-world/Julia import ordering and 47 unrelated formatting files). The test hook requires unavailable GNU timeout; its pytest command was run directly.

• Full Python 3.11 suite: 3,882 passed, 135 skipped, 66 deselected; 1 unrelated failure. Harbor port-binding test expects "already in use" but gets a startup timeout on macOS; the identical failure reproduces independently on unchanged main (cad76b7a).

Merge Danger

Door: two-way

Blast Radius: QED

Public imports remain available, but application dependencies now load when their exported classes are requested. No API signatures or architecture change; no RFC needed. Independent review and CI required before merge.

[cursor]

Alignment Review Report

Automated Checks

• Lint: FAIL (minor) — usort wants from typing import Any, TYPE_CHECKING (alphabetical) in both touched __init__.py files; they currently read TYPE_CHECKING, Any. ruff format/ruff check are clean. Note CI's lint job (test.yml) only runs over src/ and tests/, so this won't fail CI, but it's a one-line fix for consistency with the rest of the repo (e.g. atari_env/client.py, openspiel_env/client.py, sync_client.py all sort Any before TYPE_CHECKING).
• Debug code: CLEAN — no prints/breakpoints/TODOs introduced by this diff.

Open RFCs Context

Scanned rfcs/000–012. None reference QED-Math, math_verify, process pools, or lazy imports, so there's no overlap with any open RFC (several are Draft/In Review, e.g. 010 echo-env-token-world-model and 011 ard-catalog-discovery, but they're unrelated to this change).

Tier 1: Fixes Required

• envs/qed_math_env/__init__.py:10 — usort import-order: from typing import TYPE_CHECKING, Any → from typing import Any, TYPE_CHECKING.
• envs/qed_math_env/server/__init__.py:9 — same usort import-order fix.

Tier 2: Alignment Discussion

Principle Conflicts

None identified. The __getattr__/TYPE_CHECKING lazy-export pattern this PR adds to qed_math_env/__init__.py and qed_math_env/server/__init__.py matches the existing convention already used in browsergym_env/__init__.py, thinkingbox_env/__init__.py, calendar_env/__init__.py, and finqa_env/server/__init__.py (same globals() caching + __dir__ shape), so it's consistent with prior art rather than introducing a new pattern. It doesn't touch the client/server boundary, reward computation, or the Gym-like/MCP dual API, so no invariant is at risk.

RFC Conflicts

None identified.

Verification

• New tests pass: tests/envs/test_qed_math_worker_imports.py::test_spawned_verifier_does_not_import_application and ::test_public_package_exports_keep_their_identity (2/2 passed).
• Existing suite: tests/envs/test_qed_math_environment.py — 147 passed; the 5 failures are the pre-existing @pytest.mark.integration tests that need a live server on localhost:8000 (excluded by CI's -m "not integration and not network and not docker"), unrelated to this change.

Summary

• 2 mechanical issues to fix (import-sort order, non-blocking for CI but easy to clean up)
• 0 alignment points for human review
• 0 RFC conflicts to discuss

Nice, well-scoped fix — the lazy-export pattern is well-tested and matches the convention already used elsewhere in the repo.

  

_{Sent by Cursor Automation: Pre-review}

[burtenshaw]

Addressed review 5656435359 in 04d3172: reordered typing imports in the two QED init.py files. Installed usort check, ruff format --check, and ruff check all pass (no formatter disagreement). QED tests: 149 passed, 5 integration tests deselected, including spawned-worker isolation and public-export identity regressions. No main merge needed.