[Clear1nsight]

Change and mathematical scope

The actual common runner in #6864 can reap its direct child before its final numeric process-group signal. #6866 supplies a counterexample and retained-child construction. This contribution installs that construction into the connected runner and rebuilds both immutable stage recipes and their caller/source pins.

The consequential Pro review also found that the actual enclosing science.lock context could mask the working exception with a close error. The corrected actual main retains separate preallocated working/close errors, closes a known handle once, preserves the working error, and delays expected administrative exit 1 until lock retirement succeeds. A close failure remains visible to the existing stderr-sensitive refusal predicate.

The completed contribution is an inert source archive under research/h25-retained-child-lock-repair/, with actual corrected owner source, literal predecessor inverses, selected source/profile mapping, both complete stage declarations, new finite controls and raw outcomes. Scientific/native/compiler bodies and independent validators remain unchanged. This supplies no complete graph acceptance, authentic runtime admission, all-width theorem, full baseline equality or performance result.

Task and overlap

Original investigation: #6321. Maintained specification, PR map and contributor handoff.

Base: 7c49fe32c55de02c2c873655c804b16e4c3fef0a. Immediate source: #6864 at c8f7bd582ffc1301b851dde2364c60f1636aac8b. Adopted construction: #6866 at fcba647106f3e2f08de1b39ebed04a13c0a6ef8b. Conditional native/compiler argument: #6830 at 39fcd67c991bd72918080b1d2c02feb3d679206d.

Reuse #6866 instead of inventing another group owner. Retain #6864 factory/cwd and predecessor timer repairs. Adapt the primary-preserving cleanup discipline from #6339, respecting #6346's allocation/snapshot caveats. Keep #6837/#6847's primary-baseline transfer and #6762's successful-CLOSE adoption as separate open obligations.

Bounded all-author/all-state PR-body searches covered WNOWAIT, run_group_capture and post-reap behavior; the implicated heads and actual cleanup/queue sources were inspected. An unrelated merged Windows Git stdin repair was excluded by its contract. Main was refreshed before publication. This is a proportional overlap check, not an exhaustive novelty claim. Scope is refined from the public construction to its connected implementation and the actual enclosing owner.

Validation

• Fifteen new methods passed once in normal mode (0.327 s) and optimized mode (0.306 s), exercising actual selected group/predecessor source against labelled inert ports. They expose the old identity-reuse signal and verify retained ownership, once-only consuming wait, error order, capture handling, uncertainty and bindings.
• Eight distinct new methods passed once in normal mode (0.547 s) and optimized mode (0.525 s). They reproduce the old outer context's error masking; check the corrected actual lock-owner/finalizer statements using an inert working-body port; retain all six errors; check administrative-exit ordering and the actual stderr-sensitive queue predicate; reject stale runner/both stage pins and changed scientific/independent-validator inventory; and verify literal inverses and the full changed graph.
• The second operation verifies the group function is byte-identical to the first tested successor. The unchanged fifteen-method suite was not rerun.
• Each operation used clean Python 3.12, a 20-second timeout, two-second kill tail, 256 MiB AS cap and fixed output limits, with a nonrefundable 26-policy-second debit. These are fixture durations, not speedups. No actual Popen, signal, flock, guardian, compiler/native target, financial context or scientific workload ran. Return-code fixtures are not the original full optimized-refusal execution.
• Passive sha256sum -c SHA256SUMS passed for all 92 listed files; SHA256SUMS is the 93rd export file. The source map covers 61 roles and seven profiles on 66 distinct leaves. Both stage declarations retain 197-member inventories; unchanged scientific member bodies and 44 historical control references remain at exact predecessor paths. The 2,115,419-byte archive records the original absolute subject paths and is not a portable executable installation.

Pro reviewed the public predecessor/construction and identified the outer correction; it did not review the unpublished complete successor graph. Codex reviewed the exact correction, source inverse, dependent pins and recorded controls. Independent complete-graph acceptance remains open. No scientific replay hook is registered; hosted planning and skipped replay will be reported at the actual head. Physical preparation/Pro/review/reporting/publication/final-metadata costs and historical gaps remain OPEN_NOT_ZERO.

Credit

Chris R / Clear1nsight; original CrocSwap authors and the #6321/#6864/#6866/#6830 predecessors; primary-preserving cleanup #6339 and allocation/snapshot analysis #6346, with substantial OpenAI Codex and ChatGPT Pro assistance. Preserve original source notices and licences. AI reasoning is distinct from independent human or proof-kernel acceptance.

Open next questions

Independently accept the exact changed runner, enclosing lock, both recipes and complete source/helper/driver/native closure. Verify fresh-exec/default-SIGCHLD, no-auto-reap, exclusive wait and strong-reference premises. The local blocking wait after signal failure has no proved local completion bound; existing outer ownership remains necessary. In-process custody is not durable after runner death, and descendant emptiness remains unknown.

Then connect the genuine primary baseline/full comparison from #6837/#6847 and successful complete-outcome CLOSE from #6762 through the actual caller. Current native/kernel/compiler/configuration, ownership, resource/funding admission and complete normal/optimized outcomes remain prerequisites. Begin from SOURCE-MAP.json, PRO-REVIEW.md and the two literal source inverses; do not replace missing genuine inputs with fixture success records.

[Clear1nsight]

Delivered the connected retirement correction in #6867, exact head 319ea1c5f4097ef87cc0b037327f26374e93e54d, as a completed Ready for review source contribution with maintainer edits disabled.

The actual common runner retains its direct child through final group signaling and one consuming wait. The actual enclosing lock preserves the working error and exposes a close failure before administrative exit 1. Both recipes and caller/source pins select that successor. Fifteen group methods plus eight distinct lock/binding methods passed once per Python mode; unchanged controls were not rerun. All 93 published UTF-8 file bodies and Git blob identities match the frozen 2,115,419-byte export.

Exact-head hosted verification: run 42438937510, attempt 1: plan and onboarding success; replay skipped. Planning/skipped replay establishes no candidate replay or scientific qualification. Independent complete-graph, authentic caller/native/environment, baseline equality, resource/funding and complete-outcome CLOSE remain open. Complete physical costs remain OPEN_NOT_ZERO.

Maintained end-goal specification, related PRs and contributor tasks. Predecessors #6864 and #6866 remain credited; the original completion score is unchanged.

[Clear1nsight]

The complete source-binding argument for this exact implementation head 319ea1c5f4097ef87cc0b037327f26374e93e54d is now published in the reconciliation audit, at 1144f0a715e84b3f08ebe7d0e8ed8a06543f2fad. It accepts the original source-binding scope under the genuine prospective premises; it does not supply current-host/native qualification, full controls, fresh primary comparison or successful CLOSE. The maintained specification records the remaining distinct source and runtime obligations.

[Clear1nsight]

The accepted internal source argument now has two further exact-subject decisions: #6870 freezes the complete compatible all-six selection, and #6871 establishes genuine cache/full-caller and automatic-retention consumer bindings. Both retain the unchanged #6867 implementation and reuse #6869's argument. #6872 preserves the separate installation obstruction: two direct-literal launch cycles and the missing authenticated external entry/final-pin transport. The latter remains an open source acceptance condition; a construction consultation is in progress. No runtime qualification or comparison gain follows. All three scoped PRs are Ready; exact file identities and hosted planning success/replay skipped are verified. Current original-investigation overview, evidence boundaries and contribution tasks. No new guarded operation or budget.

[Clear1nsight]

The complete outer caller and native-retirement source is now published in PR #6873, exact head ae7add49484492b49b0141445fbfd6002858da54. This PR's selected interior at 319ea1c5f4097ef87cc0b037327f26374e93e54d is preserved byte for byte: all 66 selected leaves, detached manifest, both full recipes, scientific drivers and independent oracles are unchanged.

The successor adds actual authenticated parent dispatch, late-pin derivation/publication and a surviving descendant owner before G birth. Its complete source-composition/installation argument has separate challenging AI acceptance; #6869/#6870/#6871 retain their exact internal scopes. #6872's literal-launch finding and the rejected group-only cleanup remain historical evidence. The updated verification overview maps the complete lineage and contributor tasks. Exact remote file readback and planning/onboarding passed; scientific replay was skipped. Actual runtime/native qualification, resource fit, complete controls, original-primary comparison and successful financial CLOSE remain open. No new guarded operation or budget.