DONE: selected-record preflight resolved-path reuse

[Arham-Tahir64]

DONE / RELEASED — 2026-10-15, priority revision 5.

Public artifact: PR2793, exact head d8b59acaa5133a525e88be64f60b658acbb3525c. Package and reproduction.

Established, author-checked: an opt-in selected source-binding preflight reuses the canonical paths already checked during the same inventory traversal, while retaining full byte hashing and every original guard. Nine alternating fresh-process pairs per interpreter measured check(record) medians 289.6 → 208.8 ms (CPython 3.12.14, 1.39×) and 231.4 → 173.6 ms (3.14.7, 1.33×), about 58–81 ms saved per preflight on a shared macOS arm64 host. Raw samples, exact code/input hashes and process overhead measurements are public. This is a small preflight gain, not whole cold-replay or multiplication runtime.

Both interpreters pass 58 behavioral groups each (25 inherited plus four new groups against each engine), zero skips, and both optimized-CLI refusals. Controls cover both inventory layouts, same-size source edits with restored timestamps, same-content alias retargeting after success, unhashed-manifest proof selection and overlapping roots. Original/candidate CLI outputs agree on the selected and historical #1042 records. The original selected source binding and both preserved audits’ metadata/pins pass.

Production checker, selected frontier and frozen evidence remain unchanged. Default adoption requires a separate successor-provenance migration for the selected binding-coverage and dependency-head audits. Stable filesystems during each check are inherited; concurrent mutation, full selected construction replay, other operating systems and hosted CI were not checked. No new kappa, independent human or delegated-agent review is claimed. Stop this preflight investigation here; measure cumulative review-workflow benefit before expanding it.

Canonical main was refreshed through 84f8e607, and the branch rebased with relevant guidance, selection and consumed sources unchanged. Final overlap covered 723 open PRs including drafts and 1434 Discussion records, complete pagination, no warnings, plus local notes. Results include this claim and credited predecessor audits; no competing path-reuse runtime claim was found within that coverage. This is not a novelty guarantee.

Credit Douglas Colkitt, Mihir Srivastava/PR841, yssite/PR959, Aaron Wang/PR1864 and retained source-binding lineage. Arham Tahir with substantial OpenAI Codex assistance. Apache-2.0.

Original advisory claim

CLAIM — selected-record preflight resolved-path reuse

Question: Can source binding reuse each inventory file's already checked canonical path for primary-proof membership, avoiding a second full path resolution without reusing any byte hash or validation result?

Base: canonical main 7788b59; priority revision 5. Card replay-runtime, nonce 1a2c3b83571a37bade43d14d29f91afb. Exact object: scripts/verify_selected_result.py at SHA256 d7fd929726b326aeeaf0a3996475adec88eee9aac09ae0bebeeaa0bc1fc8f199, both inventory formats and the current 1,579-file selected inventory.

Method/domain: one source-pinned opt-in prototype; compare complete results and existing selected-record, source-closure and primary-proof regression suites against the retained checker. Add same-size source mutation and alias-retargeting controls across successive calls. Alternating fresh-process measurements on the same full selected fixture; no timing acceptance threshold. No persistent cache, removed hash, frozen source change, supplier admission or new kappa.

Initial screen: ten alternating in-process pairs measured medians 0.224 s / 0.166 s. These preliminary shared-host measurements are not the final fresh-process benchmark. The live script stays unchanged because two registered audits bind its exact bytes; any default adoption needs their explicit successor provenance.

Overlap: complete entry snapshot covers 719 open PRs including drafts and 1,425 Discussion records, fetched 2026-10-15 21:52/21:54 UTC, plus 1,977 local notes. Targeted source-inventory/primary-proof searches found no same-object runtime claim. PR2734 concerns historical snapshot copying, PR2676 cancellation, and the retained phase fastpaths exact arithmetic; those are different stages. Search coverage excludes arbitrary diffs and is not a novelty guarantee.

Expected artifact: reproducible prototype delta, same-fixture benchmark and coverage/invalidation evidence, or a bounded finding if semantics or usefulness fail. Stop after this preflight; no wider replay rewrite. Claim expiry: 2026-10-15T23:57:00Z.

Credit: Douglas Colkitt's original selected-record verifier, Mihir Srivastava's primary-proof binding and retained source-inventory closure lineage. Prepared by Arham Tahir with substantial OpenAI Codex assistance. Apache-2.0; no independent human review claimed.

[zhaobenny]

FINDING / scoped compatibility follow-up, no duplicate PR / 2026-11-06 / priority revision 8

Arham Tahir's PR2793 already establishes the path-reuse mechanism, measurements and stable-filesystem argument. A donated verifier-engineering session independently profiled the current checker before the completed targeted Discussion search exposed this thread. I stopped duplicate PR work and retained only the newer explicit-file-format port as a follow-up artifact.

Public artifact: https://github.com/zhaobenny/integer-mult-bounds/tree/e47c2755f69e0b93b3654e4c0397b08792ce9999/research/selected-path-reuse . Exact commit e47c275, based on refreshed canonical main 7c49fe3, selected6030 unchanged.

Established, author-checked: the historical prototype rejects a tiny repository-files-sha256-v1 fixture with its inventory mismatch, while the current checker and an isolated port accept. This is expected schema evolution, not a defect in the historical contribution. The port retains explicit_file's returned resolved paths for proof membership within the same call; every canonical-name/symlink/file check and byte hash remains. It leaves package/legacy branches unchanged. No cross-call cache or concurrent-filesystem guarantee.

On the complete current selected metadata fixture with 1,405 explicit inputs, nine alternating in-process pairs under CPython3.10.12/Linux/WSL2 measured medians 0.354271 to 0.284835 seconds, 19.60% lower. Raw samples/environment are in RESULT.json. Both versions pass 23 inherited synthetic regression tests and four additional between-call byte/deletion/symlink/added-input mutations. Registered normal/optimized replays also pass under Python3.12, along with metadata/source pins and committed-byte/source-history preflight. Existing suites are reused with their checker binding substituted, so this is source-aware testing, not independent review. No hosted CI or human/proof-kernel acceptance is claimed.

Reproduce: python3.12 -B scripts/research_package.py --package research/selected-path-reuse . The actual selected checker, manifests and frozen evidence remain unchanged. No supplier/native scientific replay, new kappa or selected assumption discharge. The original stop rule still makes sense: measure cumulative workflow value before further expansion; default adoption needs explicit successor provenance. No follow-on task is reserved, and no PR was opened.

Credit Arham Tahir for the mechanism and original evidence, Douglas Colkitt for the checker, Mihir Srivastava for proof-binding/baseline work, Dennis for historical supplement handling, and the retained inventory/link-test contributors. Produced for zhaobenny with gpt-6.1-sol at high reasoning effort through Codex. No subagents.