[Clear1nsight]

Change and mathematical scope

The common runner selected by #6864 can reap its direct child and then signal the child's old process-group number from finally. If that number is reused while the runner is descheduled, the signal can target an unrelated group. A held pidfd does not reserve the numeric namespace lookup used by killpg.

This focused source audit records the counterexample under the selected Linux v6.12 model, the exact affected runner, primary-source pins, a proposed connected correction and its required discriminating controls. The package is research/h25-group-retirement-audit. It preserves the current native launch flag and the existing factory/cwd and timer/error corrections. The proposed correction remains unimplemented and unexecuted. No theorem, selected frontier, measured gain or runtime qualification changes.

Task and overlap

Original investigation: #6321. Base: 7c49fe32c55de02c2c873655c804b16e4c3fef0a. Affected #6864 head: c8f7bd582ffc1301b851dde2364c60f1636aac8b; prior connected assembly #6863: c009a7514790e865b664ab8cd4688d37787d7eaa.

The #6830 runner at 39fcd67c991bd72918080b1d2c02feb3d679206d excludes its final kill after consumed/committed wait states. Its ownership discipline is a concrete repair predecessor, but its relay, funding and command interfaces are not a drop-in common-runner replacement. #6837/#6847 retain the separate genuine predecessor-baseline adoption obligation; #6762 retains the complete-outcome CLOSE obligation.

Bounded all-author/all-state behavior searches included killpg, start_new_session, WNOWAIT and post-reap. The inspected merged #4686 at de93ebd044d0b00a8647a9f47f86939a6b18b9aa fixes Windows Git stdin errors, a different contract. Exact dispositions and search limits are retained. No global novelty is claimed.

Scope is refined from a constructive preservation inquiry to a demonstrated negative source result and explicit repair handoff. Implementation dispatch stopped at the fixed thirty-minute reassessment before the complete Pro response arrived; publication does not reopen that round.

Validation

Codex checked the complete Pro answer against the exact affected runner and the selected Linux v6.12 PID allocation/removal, numeric signal lookup and non-consuming wait source paths. The explicit wait-before-final-signal trace establishes the defect without a newly selected CPython patch release. Official CPython v3.12.3 polling is corroborating source only.

No project code, regression, kernel/native operation, guardian, compiler or scientific workload was run. The proposed tests are acceptance checks, not reported passes. This ten-file archive has a raw-byte SHA256/Git-blob manifest and no replay hook; hosted planning and skipped replay carry no scientific credit. Complete helper/context coverage, actual source/environment association, full comparison, resource fit, retention/CLOSE and physical costs remain unresolved. Outside and physical costs remain OPEN_NOT_ZERO.

Credit

Tekalign, with substantial OpenAI Codex and ChatGPT Pro assistance. Predecessor authors and source-specific notices are preserved. Linux and CPython are referenced under their existing licenses; Linux source bodies are not copied or relicensed here. This is bounded AI source review, not human or proof-kernel acceptance.

Open next questions

Implement a single connected successor: non-consuming terminal observation, final numeric signal while the child remains owned and unreaped, then consuming wait; preserve actual first/cleanup errors. Update the authentic caller's runner pin, both complete source recipes and detached bindings together, then run the distinct reuse/ownership/error controls and obtain exact-source review. Keep the genuine primary baseline on that same common owner when it is later adopted.

The maintained goal, PR map and contributor handoff records the wider obligations. A new research window, execution allowance, successful comparison or original completion is not granted by this publication.

[Clear1nsight]

Published this scoped source audit at fcba647106f3e2f08de1b39ebed04a13c0a6ef8b. All ten committed file bodies and Git blob identities were read back against the frozen export.

Hosted checks at that exact head: plan success, replay skipped, run https://github.com/CrocSwap/integer-mult-bounds/actions/runs/42433090811. Planning/onboarding and skipped replay provide no native or scientific validation. No project control or workload was run for this source audit.

The maintained handoff links the original investigation, explains this finding's role and identifies the connected retirement correction as the next task. The proposed correction remains unimplemented/unexecuted; whole source qualification, baseline equality, actual runtime/resource fit and successful retention/CLOSE remain open. Existing predecessor repairs and their credit are preserved. Source work remains closed and outside/physical costs remain OPEN_NOT_ZERO.

[Clear1nsight]

The connected implementation following this source audit is now submitted as PR #6867, exact head 319ea1c5f4097ef87cc0b037327f26374e93e54d.

It installs this retained-child construction in the actual common runner, preserves the working error across actual enclosing-lock cleanup, and rebuilds both recipes and dependent caller/source pins. The administrative exit 1 follows known lock close. Fifteen group methods plus eight distinct new lock/binding methods passed once in each Python mode; no unchanged suite or actual scientific/native workload was rerun.

Exact-head hosted result: run 42438937510, attempt 1: plan and onboarding success; replay skipped. This completes the scoped source correction; it does not accept the complete graph, current runtime/native association, full baseline equality, resources/funding or successful CLOSE. The conditional identity premises and cross-process custody/completion gaps remain explicit. Current specification and next independent-review task.