[Clear1nsight]

Change and mathematical scope

This adds a compact source archive of a complete local passive-guard successor, its selected helper entry and source-correspondent finite fake controls. Independent conditional source/plan and publication-adapter reviews are retained; parent-run normal and optimized controls each passed all 14 cases, once per mode. Actual C10 and complete original exits remain HOLD.

The local method additionally assumes adequate capture-call and recording/unwind availability after handler delivery; it does not discharge original inline first-slot requirements. Unadmitted main still disarms the timer after initial gate failure, is not certified side-effect-free and must not be invoked.

The proposed package retains one complete successor of the failed passive guard and its selected helper entry, the exact accepted AST-extracted fake-token harness, independent scope/source review, finite fake-control receipts, and the rejected guard as inert source. Its intended result is conditional source-and-fake acceptance of the selected in-process raw-error path: preserve the first delivered exception object and rooted raw traceback, retain ordered raw secondaries, attempt every known independent close once, and stop after cleanup or acquisition uncertainty.

The accepted callable definitions and all F01–F14 finite cases are bound to the final adapted hashes in the README, adapter diff/review and raw receipts. Three stronger allocation/snapshot/recording RED witnesses remain RED. Independent raw-result review and publication metadata checks were pending at authoring freeze. They subsequently completed before submission: the independent local final review accepts the conditional source/14-case fake scope only (FINAL-REVIEW-v1.json SHA256 ba73c703f2f201de1885f361eef0a73b8e1451ba46ae3a6ff8677b95e79992b0, retained locally outside the frozen archive); archive integrity and all 35 committed-byte/mode comparisons passed. The frozen package preserves its earlier timing labels. This PR description records the later evidence; it does not turn local review into upstream scientific acceptance.

This is a source archive and finite-control record. Actual guard/runtime/native/decoder execution, original C05, full V03, actual parent teardown authority and C07 remain unqualified. Original progress stays 43.5%, 22/49 items, 2/10 stages, 1/10 milestones, 24 components and 52 helpers; the package earns zero original-completion credit. Physical and unaccounted costs remain OPEN_NOT_ZERO.

Task and overlap

Base: 7c49fe32c55de02c2c873655c804b16e4c3fef0a (canonical main refreshed by the parent this round, unchanged; exact guidance read at that pin). Explicit assignment: W1/RUN-01–03 of the bounded 2026-11-03 Pro work items; no task picker was run.

Scope is refined to the exact selected source and fake-token paths. The original actual evidence-path obligation is retained as incomplete because the external caller/admission/custody/currentness/teardown and C07 instance are absent. No new mathematical bound or measured performance is proposed.

Nearest dependencies and overlap:

• unpaidmoon PR #6339, 608ad23adc6c01264c8195daf7f18e29c17464a7: adapt primary-preserving close cleanup; retain the bare-finally failure control and successful-body close-failure condition.
• PR #6346, abff67b31fa781e950c82d95eccb6f168950943c: adapt pre-acquisition raw state and independent release/retire-before-close; preserve allocation/snapshot/recorder RED witnesses and their narrower relevance.
• PR #6347, 59140439e7f791876159e2f3838f1dead2e54953, and PR #6351, 1e9db2d08be4551ffe0bcfc573574851530eb4ca: retain the C10/C11 delivered-error premises and exclusions.
• PR #6434, edc2820ba023b804ef1351a88e2fe31bf8d489dd: formatted records/OperationStop remain incompatible with the same-object C10 oracle.
• PR #6476, 44749c3a6ddc8cde48c5aa83073f6794e3546b00: retain the accepted narrow candidate-locator semantics while adapting only the selected entry/error/close route if independently accepted.

The existing reconciliation is retained with the accepted payload. The added value is correspondence to the complete selected failed-guard successor and its caller, rather than another detached cleanup helper. The bounded overlap observation is the parent's 2026-11-03T04:38:46–48Z head refresh, not a novelty claim.

Validation

Published fork branch: Clear1nsight:codex/h25-passive-guard-primary-preservation-20261103. Remote head 928f2a997b9e8a060569f36a70b89fc7658befee, root tree 005cc8f8947371360677072d9be6119ae44cc07d, sole parent 7c49fe32c55de02c2c873655c804b16e4c3fef0a. The remote and locally reviewed commits have the same exact root tree and parent; their commit IDs differ because Git Data serialization changed commit metadata. Canonical main remains at the same base at submission.

Observed commands: /usr/bin/python3.12 -I -S -B <exact package snapshot>/verify_guard_fake_v1.py and the same command with -O, once each, on Python 3.12/POSIX. Each returned PASS_BOUNDED_FAKE_ONLY for F01–F14, exit 0, empty stderr, with exact source/helper hashes. The raw stdout/stderr/time, admission, plan and ledger are retained. No control execution, candidate import/compilation, native operation or original recipe case was performed by the publication worker.

The package deliberately has no VERIFY.json replay hook. After final freeze, python3 -I -B research/h25-passive-guard-primary-preservation/integrity.py is an integrity-only command: bounded file inventory, path/size and SHA256 comparison, no candidate import or invocation. The parent already ran that integrity-only command once on the final package and recorded PASS_ARCHIVE_INTEGRITY_ONLY: 34 payload files plus MANIFEST.json (35 files total), with the frozen manifest SHA256 83398b9c14b2b153602dc1e9f7fa86ac01fb22698d8a85e7b6f9c8685c640dc8. All 35 committed files matched the frozen bytes and regular-file modes. For this submission, a read-only remote/local hash comparison again confirmed the same 35 files (372,277 bytes); no controls or integrity runner were rerun. Source bytes, finite fake results, raw traceback/error-object checks, committed-byte identity and remote hosted status are distinct evidence.

Not run: actual guard/helper CLI, process/native/procfs/namespace paths, decoder/compiler query, original C05 or full V03 cases, historical full suites, archive exporter, cold source-history replay or performance measurement. C10 private exact-list/native/no-concurrent-async-tracing-hostile-finalizer/adequate-delivery-unwind premises and all C11 inherited/pre-handler/whole-handler-allocation exclusions remain explicit.

Credit

Prepared for Clear1nsight with substantial OpenAI Codex assistance. Retain unpaidmoon's PR #6339 authorship and source notices, Clear1nsight's PR #6346/#6347/#6351 and existing guard/helper work, the #6434 and #6476 predecessor records, all credited inherited sources and Apache-2.0 terms. Adaptation and review do not claim authorship or approval by those predecessors. Applicable retained license and notices accompany the exact accepted source bytes; publication-only in-file modification notices and dependent pins are isolated by ADAPTER.diff and normalized-AST evidence.

Open next questions

The package README retains structured questions with relevance, first step and blockers. The smallest original-scope dependency is an exact original-compatible external caller/admission/teardown and C07 instance, followed by independent review at that bound instance. Source/fake acceptance alone supplies none of that authority. No further execution or research round is authorized by publishing the archive.

The archived finite-run ledger retains canonical policy usage of 94.550533479 seconds and supplemental combined usage of 106.890533479 seconds after the two finite runs. The later separately retained metadata ledger records supplemental combined usage of 109.890533479 seconds; canonical usage remains 94.550533479 seconds. These are recorded policy debits, not complete physical cost. Preparation, review, network, reporting, publication and final-tail gaps remain OPEN_NOT_ZERO; this submission grants no new scientific budget.

[Clear1nsight]

Delivery verified at the submitted head

Chrome submitted this draft once at 928f2a997b9e8a060569f36a70b89fc7658befee (root tree 005cc8f8947371360677072d9be6119ae44cc07d, sole parent 7c49fe32c55de02c2c873655c804b16e4c3fef0a). The title/body, upstream base, fork branch, one commit, all 35 added file blobs/modes and disabled maintainer edits were read back and match the reviewed frozen package. The PR is attached to the contributing Codex chat.

Hosted checks: run 41584602277, attempt 1, completed SUCCESS on that exact head. Plan/onboarding SUCCESS; replay SKIPPED. This archive is intentionally unregistered, with no candidate replay hook. This is archival planning evidence, not guard/native/scientific execution or full qualification.

The existing canonical seven-field handoff and hub pointer were updated in place. Prior #6476 discovery work and its separate control evidence remain credited.

No unchanged controls, integrity runner, guard/helper CLI or original-data work were rerun for publication. The 14 normal/optimized fake cases and their raw-result review are prior closed-window evidence. The added capture-call/recording premise does not discharge original inline C10; all three stronger RED witnesses, the unqualified actual caller/teardown and concrete C07 lifetime gap remain explicit. Full C05/V03 remain false; original progress remains 43.5% (22/49 activities), separately 2/10 full-stage gates and 1/10 overall milestones, with zero new original activity credit.

Recorded policy usage is canonical 94.550533479 seconds and latest supplemental combined 109.890533479 seconds; complete physical and publication/final-tail costs remain OPEN_NOT_ZERO. The separate goal service was observed paused. Publication does not reopen research, resume the goal, grant a new budget or authorize merge.

Prepared for Clear1nsight with substantial OpenAI Codex assistance; predecessor attribution and Apache-2.0 notices are retained.