Stop owned POSIX descendants after direct script completion

[roger-sato]

Summary

When a script starts an ordinary POSIX solver with redirected output and exits, run_parallel.py can print PASS all 1 scripts and return while the solver continues running. A failed leader leaves the same leak. Clean up the owned process group before unregistering the completed leader, serialized with the controller's cancellation snapshot. Preserve successful/failed output, exclude unrelated groups, and fail visibly if cleanup fails.

Add five actual-CLI regressions and run them in the existing Ubuntu 3.12 reconciliation workflow. Extend its historical adapter with only the exact new runner hash; the frozen runner, reconciliation manifest and source pins remain unchanged. The new note records reproduction, scope and validation.

Based on canonical main 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority revision 6. Builds on ywz65's #2676 cancellation correction, its independent #2777 audit, and the isolated-test completed-leader cleanup. This covers normal leader completion; it makes no native Windows, detached-group, execution-deadline or mathematical claim.

Test plan

Local environment: macOS arm64, CPython 3.12.14, Git 2.54.0.

• python3.12 -W error -B -m unittest discover -s tests -p 'test_script_exit_cleanup.py' -v: all 5 pass; the original runner fails 3. Covers children/grandchildren, unrelated-process survival, success/failure output and an already-joined group. Parent-only cleanup mutation is detected; injected signal denial rejects success and preserves retry diagnostics.
• python3.12 -O -W error -B -m unittest discover -s tests -p 'test_script_exit_cleanup.py' -v: all 5 pass.
• Exact added workflow command, python3.12 -W error -B -m unittest discover -s tests -p 'test_script_*.py' -v: 14 tests, 5 pass and 9 existing Linux/Windows observers skip on macOS.
• Existing cancellation/parallel/assertion/memory modules: 30 tests, 20 pass and 10 platform skips; exact bounded suite command is in the note. Independent thread/lock race controls verify both cleanup ownership orders stop once and retain the original exception.
• Actual python3.12 -B scripts/replay_historical_check.py reconciliation92 in its complete 383-file closure passes 92 pinned trees / 4,564 distinct blobs. Unknown live runner and corrupted frozen input both reject; original manifest and sources stay exact.
• python3.12 -B scripts/research_package.py --package research/isolated-test-bytecode and --package research/isolated-descendant-bytecode: every registered command passes; declared outcomes remain completed / partial.
• git diff --check.
• python3.12 -B scripts/research_package.py --package research/historical-snapshot-one-pass has an unchanged baseline failure: 27 tests, 1 failure + 1 error, both from the unrelated tooling-ledger guard for docs/ci-verification.md. The same result is reproduced with the actual registered CLI over 82 files read exactly from canonical main. No unrelated hash admission was added. Fork-hosted results are recorded below; upstream approval is separate.

Hosted verification reviewed on 2026-10-18

• Fork-hosted Ubuntu 24.04.5 / CPython 3.12.15 replay at exact commit 58a754bf08ce63080822a28451aaa4952e2f6e2a: reconciliation92 workflow passes 92 pinned trees / 4,564 distinct blobs, all 6 worker-isolation tests and all 14 direct-script tests (including all 5 new regressions, no platform skips). The passing test process emitted ignored resource-finalization warnings in the existing cancellation controls; this is not a warning-free run.
• The separate Community exact tooling run passes tooling-exact and the scoped owner/provenance diagnostics, then stops at the known docs/ci-verification.md successor guard addressed by #3648. Subsequent historical/reconciliation tests and the tracked-diff step were not reached. This does not establish that the complete tooling workflow passes after that correction.

These are completed fork workflow results, not upstream maintainer approval or a mathematical replay.

Prepared for roger-sato with substantial OpenAI Codex assistance; independently reviewed with AI assistance. Apache-2.0.