[yaoyu-33]

Release the Python worker through the shared resources lifecycle so close works independently of verification. Switch worker start from the default multiprocessing context (fork on Linux) to a lazily started forkserver worker, keep request-based scoring and serialize allocation/cleanup.

Depends on #5709 (which depends on #5708). The child diff contains only this environment and its related documentation/configuration; shared implementation is reviewed once in the dependency. No separate issue is needed for this focused migration of an existing resources server.

• Track the shielded worker-start task until its result is published, including cancellation of the request awaiting it. Close waits for pending allocation and then releases the actual worker; an abandoned to_thread result can no longer leak a process.
• Add forkserver socket isolation and an async close that escalates SIGTERM then SIGKILL off the executor. A surviving process remains owned and makes cleanup fail for retry. Verification now uses the shared public context; bounded receipts retain identity, while scoring recomputes from the request.

Request-scored: retains an empty marker (retain(None)) because the request already carries the answer and grading data.

Validation on 7cc1eb7c8fece27c1302991bd4e7e748976dc3c1, Python 3.13.14, with this checkout first on PYTHONPATH and NEMO_GYM_EXTRA_ROOTS pointing to it:

• python -m pytest -o addopts='' -o log_cli=false resources_servers/math_with_code/tests -q --tb=short: 45 passed.
• The shared lifecycle and MCP checks run in feat(sessions): centralize resources lifecycle and verification ownership #5709 (1,251 passed at its head, with main merged) and fix(mcp): preserve seed validators and annotated request injection #5708 (69 passed); this PR does not touch those files.
• ruff check, ruff format --check and git diff --check passed on the changed files. The private guardwords catalog was unavailable, so that scan was skipped. New commits include DCO sign-off.

Real-model smoke on preceding migration revision 27a4d4998ba824d9dc06ac41f11a914cffff9d67 (shared core 4e92d4891) used nvidia/nemotron-3.5-lightning-30b-a3b, Python 3.13.14 and actual forkserver workers. Normal and skip-verification cases each executed Python and returned 385. Normal verification scored 1.0; the skip case made zero verify calls and returned fixed reward 0.0. Verify/close returned 200 as applicable, worker death was observed, and active/starting/identity maps were empty. The skip case released a real live worker at close. Loaded source hashes matched that revision. These are synthetic lifecycle tasks, not a math benchmark baseline. The smoke made one verify per episode, so the retry path (retain(None)) is covered by unit tests only.

These targeted checks do not replace the repository CI matrix or approval of the current head.

Compatibility

The first worker per server process waits for the forkserver to import the module (seconds; later workers start in milliseconds). The forkserver is one extra resident process per server process (about 230 MB measured on macOS). Closing a worker that ignores SIGTERM can take about 11 s while holding the session lock. A worker that cannot start still answers the tool call with success: false (as on main), is now logged, and the episode is scored from the answer alone.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

/ok to test d87a7f4

[yaoyu-33]

/ok to test af49dc9

[yaoyu-33]

/ok to test 7cc1eb7

[yaoyu-33]

/ok to test d11dc70