[burtenshaw]

Summary

Add regression coverage for QED worker import isolation and busy-worker cleanup. Main now contains the lazy-export production fix from #1491; this PR adds tests only.

worker import → excludes client, models and full environment
busy worker → timeout/cancellation → admission released → subsequent verification succeeds
stop → executor reference cleared

Retains public export identity and fresh-process import coverage. Request deadlines, rewards and lifecycle implementation are unchanged.

Evidence

• Before: main has cold-worker import tests, but lacks the explicit models exclusion and these busy-worker timeout/cancellation recovery cases.
• After: at 68060a4391573a3d024af8a779d2c3a3933d2c28, both QED test files pass: 153 passed, 5 deselected on macOS/Python 3.11. Four fresh-process import probes and scoped Ruff/usort checks pass.
• The 76 added test lines are unchanged by conflict resolution; both production initializers match main exactly.
• New-head CI is pending. Earlier Linux CPU-pressure evidence validated the production optimization now on main; it was not rerun for this test-only head. These tests do not establish termination of arbitrary running jobs or resolve Add optional fastapi-guard security middleware for env servers (default off) #1361 policy gates.

Merge Danger

Door: two-way; revert the added tests.

Blast Radius: tests. No production behavior changes.

[cursor]

Alignment Review Report

Automated Checks

• Lint: PASS — uv run usort check / ruff format --check / ruff check all pass on the three files this PR touches (envs/qed_math_env/__init__.py, envs/qed_math_env/server/__init__.py, tests/envs/test_qed_math_environment.py). (Note: a full-repo ruff format --check reports 47 files needing reformatting, but none of them are touched by this PR — pre-existing drift, not introduced here.)
• Debug code: CLEAN — no print/breakpoint/pdb/TODO/FIXME in the diff.

Verification performed

• Installed the env's own deps (math-verify, datasets, fastmcp, trackio) and ran tests/envs/test_qed_math_environment.py: 151 passed, including the two new tests (TestVerifierImports::test_worker_import_does_not_load_environment_dependencies, TestVerifierImports::test_public_exports_preserve_identity) and the new test_busy_worker_deadline_and_cancellation cases. The 5 errors in TestQEDMathServerIntegration are pre-existing (that class predates this PR) and only fail here because no live server is running in this sandbox — unrelated to this change.
• Confirmed the actual fix in a fresh interpreter: import qed_math_env.server.math_verify_service no longer pulls in qed_math_env.client, qed_math_env.models, datasets, fastmcp, or openai — exactly the problem a spawn-context ProcessPoolExecutor worker hits on re-import.
• Confirmed the lazy __getattr__/__dir__ (PEP 562) pattern preserves the public API: from qed_math_env import QEDMathEnv, qed_math_env.server.QEDMathEnvironment, and dir(qed_math_env) all behave as before; identity checks in the new test confirm the lazily-loaded objects are the same objects as direct submodule imports.
• Checked for the same risk elsewhere: no other environment under envs/ uses ProcessPoolExecutor/mp.get_context("spawn"), so this fix is correctly scoped to qed_math_env alone per the self-contained-environments guidance (no sibling copies need the same fix).

Open RFCs Context

Scanned all RFCs in rfcs/ (000, 001, 002, 003, 004, 005, 006, 008 — status In Review; 010, 011 — status Draft). None reference qed_math_env, process-pool verifier workers, or package __init__.py/lazy-import conventions, so there is no active design discussion this change intersects with.

Tier 1: Fixes Required

None — mechanical checks are clean on the changed files.

Tier 2: Alignment Discussion

Principle Conflicts

None identified. This is an internal import-ordering fix inside a single self-contained environment:

• No change to the Gym-like API signatures or MCP tool surface (Invariant: API Invariants).
• No crossing of the client/server boundary — envs/qed_math_env/__init__.py still only re-exports client/model symbols, and server/__init__.py still only re-exports the server Environment class; both via lazy attribute access rather than eager imports (Invariant: client-server separation, git-blamed to @Darktex, who authored this section of INVARIANTS.md).
• No change to reward computation or where it lives (Principle: rewards inside environment).
• Pattern is isolated to qed_math_env and doesn't change the canonical openenv init scaffold in PATTERNS.md for other environments.

RFC Conflicts

None identified.

Summary

• 0 mechanical issues to fix
• 0 alignment points for human review
• 0 RFC conflicts to discuss

This is a narrow, well-tested fix: it correctly diagnoses that ProcessPoolExecutor(mp_context=mp.get_context("spawn")) re-imports qed_math_env.server.math_verify_service in each spawned worker, which previously eagerly pulled in datasets, fastmcp, openai, and the full QEDMathEnvironment/QEDMathEnv stack on every worker start. The TYPE_CHECKING + __getattr__/__dir__ (PEP 562) lazy-export pattern keeps the public API identical while avoiding that cost, and the new tests lock in both properties (lightweight worker import, and preserved public export identity/AttributeError behavior).

  

_{Sent by Cursor Automation: Pre-review}