test: cover Linux descendant cleanup in research replay

[p-romeo]

The research replay timeout test currently checks one sleeping process and the exception. It can miss a regression that kills the immediate verifier while leaving its child solvers running. This adds a Linux-only regression test with a real parent/child/grandchild tree.

The fixture waits for both descendants before starting the timeout, verifies their PID/group/session relationship, and checks that no recorded process remains live. A normal-completion control passes; replacing group termination with parent-only termination is detected by the same assertion. Teardown retains the real group kill and the fixture has a bounded lifetime. Only tests/test_research_process_tree.py is added; the existing research workflow discovers it automatically. Production code, frozen inputs and the selected frontier are unchanged.

Evidence

• Linux/WSL2, CPython 3.12.14: python -B -m unittest discover -s tests -p 'test_research_*.py' -v — 23 passed, no skips, including all three new tests.
• Linux/WSL2, CPython 3.14.4: focused module — 3 passed.
• Windows, CPython 3.12.14: focused module — 3 explicit Linux-only skips.
• Independent compatibility check of PR Tooling: Windows portability for research onboarding scripts and tests (no new kappa) #1298 at e76026571dbfa7299bf474b941c9abeb5bd556b3: its six changed files were overlaid on 34 source files exported from main 4040a0f46d67b41c36a472a7cd4af791fd4c36cc, with each Git blob and SHA-256 checked. All 23 existing tests passed, then 26 passed with this module and -X utf8=0. The original Windows audit did not cover Linux. This is corroboration of that implementation, not another portability fix.
• Separate agent source review found no actionable issue; git diff --check passed.

Reproduce the added coverage on Linux with Python 3.12+:

python3 -B -m unittest discover -s tests -p test_research_process_tree.py -v

Scope, coordination and credit

Claim and audit: Discussion #1472. Initial main: 4040a0f46d67b41c36a472a7cd4af791fd4c36cc; submission base: e6f9716621f4e01fc190617934cdfaf072bd976f. The guidance, selected record, launcher and existing test sources did not change between these bases.

Task scope changed from the selector's bit-helper-contract suggestion (nonce 06351c3b61dce7f1b61b1a4517a5f6c9) to a portable-replay verification audit after the prescribed overlap check encountered the known Windows Unicode failure. The original card's positive control passed. Live PR-body/Discussion/README/ledger searches found #1298's fix and @woklerx's independent Windows Python 3.12 audit; this adds the missing Linux execution evidence and descendant cleanup regression. Budget/stop rule: one bounded exact-head audit; no mathematical construction or search.

Credit Douglas Colkitt for the launcher and onboarding tests, @rtyu193-rgb for #1298, and @woklerx for the prior Windows audit at a6f02272aed71a70c8107cc61bce763267fa2656. Prepared by Paul Romeo (@p-romeo) with substantial OpenAI Codex assistance in investigation, implementation, testing and writing. Apache-2.0; no independent human review is claimed.

Limits and open next questions

This checks ordinary descendants that stay in their inherited process group/session. It does not test escaped daemons, Windows taskkill, macOS, or the full selected mathematical cold replay. No kappa improvement is claimed. Hosted CI is separate from the local Linux results.

A useful next check is the same readiness-based descendant assertion for #1298's Windows branch. It would cover a behavior the current platform-specific test skips; the first step is a Windows process-state observer and parent-only failure control. That platform-specific observer is outside this PR.

[Mister-Meeseeks]

Batch review at 225cc15f88732ef86f48b75e34a9422bcb664993.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Linux-only descendant cleanup controls passed in Docker, including the deliberately broken parent-only killer; macOS's skipped run is retained separately. The tests exercise the existing bounded runner and add no mathematical acceptance rule.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.