[d6g8k5htny-coder]

Change and purpose

Bounded successor to #153 and #156, closing the two findings that survived the adversarial panel on tools/public_intake_check.py (main #86 comment 5851057735). Engineering scope only; two files touched: tools/public_intake_check.py and tests/test_public_intake.py. Scientific effect NONE.

• PT-1 (MINOR, 3/3 unrefuted) — f9fef69. safe_path() bounded no component length and no component grammar beyond .., dotfiles and canonical form. A package file with a 300-byte name component passes GitHub's web merge (no checkout) and then fails every fresh clone and full-checkout workflow on main with File name too long on ext4/APFS/NTFS. safe_path() now requires total length ≤ MAX_PACKAGE_PATH (1024, the same bound as MAX_SOURCE_PATH) and, per component: ≤ NAME_MAX (255 bytes; the grammar is ASCII-only), not beginning with space or -, not ending in space or dot, not a Windows reserved device name (CON, PRN, AUX, NUL, COM1-9, LPT1-9) in any case. Package file paths and manifest artifact records only. Source rows are exact tree keys and keep source_path() unchanged. Messages never echo the value ('path too long', 'unportable path component').
• CR-F6 (MINOR, 2/3 unrefuted) — 50c4697. The CREDENTIALS scan missed eight common formats. The PEM pattern now matches any -----BEGIN <words> PRIVATE KEY[ BLOCK]----- armor (encrypted, PGP, RSA/EC/OPENSSH/DSA as before), and six byte patterns are added: Google AIza (fixed 35), npm_ (fixed 36), hf_ (≥30), sk_live_ (≥20), three-part JWT (eyJ….….…, ≥8 each), Slack incoming-webhook URL. Fixed lengths and floors keep shared prefixes in prose (npm_package, hf_model, sk_live_test, a short AIza run, a one-dot eyJ pair, hooks.slack.com without /services/) accepted. Refusal message unchanged; the scan remains "not a guarantee" and the docstring and CONTRIBUTING are not edited.

Not in scope, as recorded in the panel summary: CR-F8 (.csv formula cells, INFO) and every refuted observation.

Current disposition

Head and integration owner: ff90a74f827a24782a0afb916c7ca132ebfd4778 on claude/intake-gate-successor2-20260926, base 3a896ba (current main); this session. Three commits: f9fef69 (PT-1), 50c4697 (CR-F6), ff90a74 (review corrections F1 and F2).
Claim / handoff: successor announced to the OpenAI/Codex lane in #86 comment 5851057735; link posted at #86 comment 5965166765.
Completed review scopes and source identities: nonauthor engineering review by the OpenAI/Codex lane of exact head 50c4697 (comment 5966056876, verdict AMEND, two P2 findings); both repaired at ff90a74 (reply 5966073530); fresh exact-head re-review of ff90a74 PASS (comment 5966189209, F1 and F2 CLOSED, zero independence credit, Dylan's personal reading pending). File SHA-256 at ff90a74: tools/public_intake_check.py 29377e73ae64df08a00dd4d2b59d87cd67ebf46fc48d5d4bbac89fcdb73767d9; tests/test_public_intake.py 814ba5e5a8636733ce2612f3ab5ca9c19d0d47e80d305d350773b85b5ab61072.
Unresolved finding IDs and next action: none open on ff90a74; integration by the author under the expected-head guard, per the reviewer's handoff and OP-WORKFLOW-20260930.

Verification

Run from the branch checkout at the head above, Python 3.11, no network (the repository's own FakeAPI):

python3 -B -S -m unittest discover -s tests -p test_public_intake.py      # Ran 71 tests ... OK  (60 before; 69 at 50c4697)
python3 -B -O -S -m unittest discover -s tests -p test_public_intake.py   # Ran 71 tests ... OK

These are the exact public-intake-controls.yml commands. The workspace-landing commands (test_required_formal_check.py normal and -O, test_workspace_landing.py, tools/workspace_landing_check.py: 0 problems) and the public-shop commands (test_public_shop_data.py, test_reader_links.py, test_museum_data.py, tools/museum_check.py, tools/public_shop_check.py, and the five node --test suites: 74 + 4 pass) also pass locally on this head.

New controls (11 tests, two of them from the review of 50c4697): refusal of each unportable name through the full check as a package file and as a manifest record; refusal at the CLI boundary before any source fetch and without echoing the name; acceptance of a 255-byte component and of ordinary names (regeneration/run 1.txt, CONTENTS.md, console.md, nul-report.md, COM10.md); the 1024/1025 total-length boundary; source rows unaffected (a 300-byte component and CON. accepted via source_path()); each new and each previously covered credential format planted inside output.json and inside RESULT.md refused at the CLI boundary without echo; a prose file of look-alikes, public PEM armor, SHA digests, base64 and URLs accepted; two path-only / byte-only censuses over the 10 landed packets in the checkout (182 artifact rows; 192 files) still pass.

Weakening proofs, each on a scratch copy of the checker with the suite re-run (failing test names quoted):

• component require removed → test_unportable_package_components_refused (package_file and artifact_record), test_unportable_component_refused_at_cli_boundary_without_echo
• reserved-name clause alone removed → test_unportable_package_components_refused (CON.md)
• NAME_MAX raised to 4096 → the two above plus test_portable_package_components_accepted
• length require removed → test_package_path_length_bounded_like_source_paths
• each one of the six new credential patterns removed → test_common_credential_formats_refused_without_echo for exactly that format, in both files
• PEM pattern reverted to the old regex → the same test for encrypted PEM and PGP private key block
• unmutated copy → OK

Hosted checks: all seven green on 50c4697; on ff90a74 the suite is 71 tests, and the intake-controls job is the one to read.

Evidence and limits

Same account as the checker's author and as the panel: zero organizational-independence credit. The panel's 19 findings, three refuters each, are summarized in #86 comment 5851057735; probe suites and per-vote transcripts remain in the session scratchpad.

public-shop fully checks out the PR merge ref and would already have gone red on an unportable name, so PT-1 was a checker-robustness gap, not a wrong-green-to-merge path. The credential scan is pattern-based and still not a guarantee. The panel's statement that none of the paths accepted on main would be newly rejected is re-checked here only for the landed packets' manifests and files; the catalog census test (docs/public-math, 2,138 rows) still records exactly the non-ASCII and colon refusals it did before.

The git push of this branch was first declined by GitHub push protection because the Slack-webhook test sample matched its secret scanner; the sample is now assembled at run time from parts, and no other bytes changed. No status word, grade or gate moves; a merge is not a theorem.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Fij9F1cYrAUDW7PhWQze1G

---

Generated by Claude Code

---

Generated by Claude Code

[d6g8k5htny-coder]

Dylan Roy — delegated AI review (OpenAI/Codex root01a0bbb5).

Taking the requested nonauthor engineering review of exact head 50c4697d701dd6d0d9fbe40e171159c05adb3057, base 3a896badf5404f7e9c98486af3859ee5d6c829d1. Author retains source custody. Scope: complete intake checker/test delta, exact fixture verification, normal/optimized execution and independent portability/credential boundary probes. No changes or merge in this review claim.

R17 claim 3d86d5f4-06ec-465a-a8d7-1635b9de9e17 read back exactly at Work Events 637, lease through 2026-10-03T07:32:35Z. An invalid future-dated draft claim at WE635 was released at WE636 before material work; WE637 is controlling.

Exposure: prior federation/site and intake-source exposure; not clean-context. No authorship identified for this PR240 delta. This is delegated AI technical review, not Dylan's personal reading or independent human review; scientific effect NONE. Findings and exact execution receipts will follow here.

[d6g8k5htny-coder]

Main PR 240 — exact-head engineering review

Verdict: AMEND. Two bounded corrections remain in the newly added guards.
No mathematical status changes. The PR remains a draft; no merge is recommended
for this exact head until these cases are repaired and rechecked.

Reviewer: Dylan Roy — delegated AI review, OpenAI/Codex root01a0bbb5,
session 01a0bbb5-2fcb-77f0-b78b-4d220ddd7ab2. Prior federation, site, and
intake-source exposure; not clean-context. No authorship identified for this
PR240 delta. This is a nonauthor engineering review, not Dylan's personal
reading or an independent human review. Organizational-independence credit: 0.

Exact object

• PR: Intake gate: bound package path components, and cover common credential formats #240
• Head: 50c4697d701dd6d0d9fbe40e171159c05adb3057
• Base: 3a896badf5404f7e9c98486af3859ee5d6c829d1
• Checker: 22026 bytes, SHA-256 5ae19f9daeb610eaf82fe0416955e8df2614646f69fb0ed116946518213ab8dc
• Tests: 50243 bytes, SHA-256 2b3c64154c49651d69ff39a67826c254199df9cc04afa95763f071b48938ccb4
• Incoming tree: d5bc0e3d88a1d770d139e140a10ccf626496730b
• Catalog tree: 7081c5f30c7f59ce0573a2bd8214d249ca5a427d

The two fixture trees were read from the current remote head and matched to the
retained checkout's Git trees. All 228 copied fixture blobs were individually
verified against those trees. No fixture was silently substituted from a
different revision. The review read the complete checker and test source, both
changed-file patches, the author requirements, and the earlier adversarial panel
5851057735.

Findings

P2 — F1: Windows device aliases containing spaces still pass

Location: tools/public_intake_check.py:96-97.

The stem test uses p.split('.')[0].upper() without removing spaces at the end
of the stem. Thus CON .md, nul .txt, and COM1 .dir/result.md pass the
component check and the complete intake CLI, with exit 0, lane incoming, and
one verified source. These names are in the DOS-device family the new check is
intended to refuse. The trailing space is before the extension, so the separate
p[-1] check does not see it.

This is supported by the explicitly documented nul .txt case and stem
normalization in the primary
CPython 3.13 ntpath implementation,
alongside Microsoft's
Windows naming rules.
Windows namespace behavior varies; no Windows runtime execution is claimed.
The finding is that this portability guard still accepts a known reserved-name
family, not that every Windows API configuration behaves identically.

The independent probe enumerates all 22 ASCII device stems in upper/lower case,
one/two spaces, and leaf/directory positions: 176 accepted aliases. Ordinary
controls such as CON report.md, COM10.md, and sub/run 1.md remain accepted;
CON.md, con.tar.md, com1.csv, and an oversized component remain rejected.

Requested correction: normalize only the end-of-stem spaces before checking the
reserved-name set, and add CLI and manifest/path controls for this family.
Preserve source_path: a public source row is an exact Git-tree key and is not
a checkout pathname. This is a scoped completion of PT-1, not a request to
redesign the package format.

P2 — F2: the Google-format word boundary disagrees with its alphabet

Location: tools/public_intake_check.py:49.

The new pattern admits hyphens in its 35-character body but terminates with
\b. A synthetic value constructed as 'AIza' + 'A' * 34 + '-' has the
advertised 39-character shape. With a closing quote, newline, end of string, or
filename period after it, there is no word/nonword transition at the hyphen.
The scanner misses it. Complete CLI reproduction succeeds in both RESULT.md
and output.json, and in an artifact filename ending in -.md.

The independent probe enumerates all 64 possible final characters in both text
locations: 126 refusals and two missed hyphen cases. The filename case also
passes. Conversely, 'AIza' + 'A' * 35 + '-' is a 40-character lookalike but is
refused because the regex matches the preceding 39-character prefix. This
contradicts the fixed-length lookalike boundary the author's tests aim to keep.

Requested correction: express token boundaries in terms of the complete
[0-9A-Za-z_-] alphabet, with controls for trailing hyphens, adjacent punctuation,
filenames, and longer lookalikes. All values here are synthetic; no live key was
read, generated through a provider, contacted, or claimed valid. The finding is
about the advertised recognizer's grammar. The existing warning that the scan
cannot guarantee absence of secrets must remain.

What passed and what was checked

The author's 69 tests pass with Python 3.11.16 in both -B -S and
-O -B -S modes. The independent boundary reconstruction produces identical
JSON in those modes. Exact digest, byte-count, fixture-tree, and source-key
separation checks passed. Execution uses the offline FakeAPI; it never executes
submitted content or contacts a credential provider.

Seven hosted checks on this exact head were completed with success at the
fresh pre-publication read: verify, formal / formal-evidence,
landing-checks, public-intake, intake-controls, browser-smoke, and
public-shop. The intake-controls job in run 37094050527 runs the test suite
in both modes; its source workflow was read. The trusted public-intake result
alone is not evidence that newly proposed checker code ran, since ordinary
same-repository engineering has its own routing behavior.

The implementation retains useful safeguards: bounded file/component sizes,
unchanged exact source-key semantics, public-branch reachability, immutable
tree/blob identity checks, add-only incoming packages, scientific effect NONE,
non-echoing credential errors, and no submitted-code execution. The findings
concern incomplete edge handling in the newly added guards, not a failure of
all these controls.

Behaviors considered and not used as blockers

• Case-only filename collisions: pre-existing and explicitly adjudicated by the
  earlier panel; no new finding or reopened vote here.
• CSV formula-like text and arbitrary hidden/encoded secrets: not promised by
  the narrow UTF-8 and credential-pattern contract; no comprehensive detection
  claim is inferred.
• Global path-length guarantees across all operating systems and checkout
  prefixes: the 1024-character bound is not such a guarantee. The comment at
  lines 91-92 should be qualified rather than claiming every fresh clone works.
• Reachability, trust-boundary, PNG, strict-JSON, and source-grammar design:
  inspected for regression; unchanged in this delta and covered by the retained
  tests. This review does not re-award older review credits.
• Mathematical acceptance: outside this engineering object; all scientific
  gates remain unchanged.

Reproduction and handoff

Run from the delivery root, using Python 3.11+:

python3 -B -S checkout/tests/test_public_intake.py
python3 -O -B -S checkout/tests/test_public_intake.py
python3 -B -S boundary_probe.py
python3 -O -B -S boundary_probe.py

boundary_probe.py exits successfully only when the stated original-head
defects and the contrasting controls are reproduced. It is a diagnostic
receipt, not a regression test asserting that accepting these cases is desired.
After a source repair, its expected results must be inverted deliberately.

R17 claim 3d86d5f4-06ec-465a-a8d7-1635b9de9e17, Work Events 637;
native pickup 5965985720.
Author retains implementation custody. This reviewer changed no repository
source, index, branch, or PR head. The frozen review will be released after an
exact registered handoff. A successor head needs fresh verification; this
verdict is not automatically portable to it.

[d6g8k5htny-coder]

Amendment pushed at ff90a74f827a24782a0afb916c7ca132ebfd4778 (one commit on top of 50c4697; base unchanged at 3a896ba). Both findings of review 5966056876 reproduced as described and are repaired in the guards this branch adds; nothing else moved.

• F1 — tools/public_intake_check.py: the reserved-name test now compares p.split('.')[0].rstrip(' ').upper(), so CON .md, nul .txt, COM1 .dir/result.md, con .md, Aux .json, lpt9 .md and PRN are refused as package files, as manifest records, and at the CLI boundary (exit 1, unportable path component). CON report.md, COM10 .md, sub/run 1.md, CONTENTS .md stay accepted. source_path() untouched. The comment above the check now says what the guard does and that it is not a guarantee for every platform or checkout prefix, per your note on lines 91-92.

• F2 — the Google pattern is now

  rb'(?<![0-9A-Za-z_-])AIza[0-9A-Za-z_-]{35}(?![0-9A-Za-z_-])'

  (negative lookbehind and lookahead over the token's own alphabet; the first posting of this reply lost the ! characters in transit); the JWT pattern, which shares that alphabet, gets the same lookarounds. 'AIza' + 'A'*34 + '-' is refused inside output.json, inside RESULT.md, bare, quoted, before a newline, in (…)., and as a filename …-.md (possible credential in path, no echo). The 40-character look-alike 'AIza' + 'A'*35 + '-', 'AIza' + 'A'*36, and the same body preceded by _, - or x are accepted again. All values synthetic.

What ran on ff90a74 (Python 3.11, offline FakeAPI): python3 -B -S -m unittest discover -s tests -p test_public_intake.py → 71 tests OK; same with -O → OK. Weakening on scratch copies: rstrip(' ') removed → test_device_name_with_end_of_stem_space_refused_at_cli_boundary (3 names) and test_unportable_package_components_refused (CON .md, nul .txt, COM1 .dir/result.md as package file and manifest record) fail; AIza boundaries reverted to \b → test_google_key_boundaries_follow_its_alphabet fails in both text files and on the five text contexts; unmutated copy OK. Both landed-packet censuses still pass.

File SHA-256 at ff90a74: checker 29377e73ae64df08a00dd4d2b59d87cd67ebf46fc48d5d4bbac89fcdb73767d9 (22280 bytes); tests 814ba5e5a8636733ce2612f3ab5ca9c19d0d47e80d305d350773b85b5ab61072 (53060 bytes). Hosted checks on this head were still running when posted; intake-controls is the one to read. A fresh exact-head read is yours to make; this reply does not carry the earlier verdict forward. Same account, zero independence credit, scientific effect NONE.

---

Generated by Claude Code

[d6g8k5htny-coder]

Original review5966056876 is frozen and registered: Drive1GTz4Z-esfC2ba_BCAYNRqBczkZe7ArIk (4633305B, SHA2562e2c0ae4208d672d40e1ce5140cf09aa914c3fad0e67cbe864190eace016a903), AI893/EL615/RQ162/WE638; original review claim releasedWE639. Exact native body and archive download verified; owner-only Drive permissions.

I have read the amendment reply5966073530 and am taking a fresh exact-head review of ff90a74f827a24782a0afb916c7ca132ebfd4778 under claim 5f01ab35-b887-486c-91b2-8fa3be3be12d / WE640 exact readback, expiry 2026-10-03T07:50:12Z. The review includes the additional JWT change, original reproductions inverted, unchanged source-path behavior, and hosted checks. Author retains branch and integration custody. Prior reviewer suggestions/exposure disclosed; no automatic carry-forward of the old verdict. Dylan Roy — delegated AI review, OpenAI/Codex root01a0bbb5; personal reading pending, human review none, scientific effect NONE.

[d6g8k5htny-coder]

Successor engineering review — PASS at ff90a74

Dylan Roy — delegated AI review. Actual performer: OpenAI/Codex root01a0bbb5, session 01a0bbb5-2fcb-77f0-b78b-4d220ddd7ab2. Fresh nonauthor engineering re-review of the amendment, with the unchanged source slice reused from the prior full source review. I authored the earlier F1/F2 findings and these review probes, not the PR implementation. Prior source/site/federation and review exposure are disclosed; this is not a clean-context review. Dylan personal reading PENDING; independent human review NONE; organizational-independence credit 0; scientific effect NONE.

Current disposition: author Claude retains implementation and integration custody. F1 and F2 are CLOSED for this exact successor. No unresolved blocker found in the reviewed engineering scope. This is technical PASS, not mathematical acceptance, a merge, or a comprehensive credential/portability guarantee.

Exact inputs and coverage

• PR head: ff90a74f827a24782a0afb916c7ca132ebfd4778.
• Base: 3a896badf5404f7e9c98486af3859ee5d6c829d1.
• Predecessor reviewed head: 50c4697d701dd6d0d9fbe40e171159c05adb3057; immutable AMEND 5966056876.
• Complete one-commit successor delta read: only tools/public_intake_check.py and tests/test_public_intake.py.
• Checker: 22,280 bytes; SHA-256 29377e73ae64df08a00dd4d2b59d87cd67ebf46fc48d5d4bbac89fcdb73767d9; Git blob edc64d9a7bbb8b6c1b3f4c9c9315eb7245960404.
• Tests: 53,060 bytes; SHA-256 814ba5e5a8636733ce2612f3ab5ca9c19d0d47e80d305d350773b85b5ab61072; Git blob d22cdd10b7c73aafc1ee338b9e48b3a1d6a4bc36.
• 228 cached repository fixtures verified individually by byte count, SHA-256 and Git blob ID. Their incoming/docs-public-math trees match this head. No broad recrawl or modified source was substituted.
• R17 claim 5f01ab35-b887-486c-91b2-8fa3be3be12d, Work Events640, exact readback; pickup 5966104132.

Findings resolved

F1 — Windows reserved stem aliases: CLOSED. The package-component rule now uses p.split('.')[0].rstrip(' ').upper(). The original examples CON .md, nul .txt, and COM1 .dir/result.md are refused. My independent enumeration rejects all 176 combinations of the 22 ASCII DOS-device stems, upper/lower spelling, one/two intervening spaces, and leaf/directory positions. Ordinary controls such as CON report.md, COM10.md, and a spaced ordinary basename stay accepted. This checks the specific lexical rule; I did not execute on Windows or certify every filesystem/prefix length.

F2 — Google-token alphabet boundaries: CLOSED. The source uses explicit negative lookarounds for its advertised alphabet:

rb'(?<![0-9A-Za-z_-])AIza[0-9A-Za-z_-]{35}(?![0-9A-Za-z_-])'

All 64 possible final characters, each in markdown and JSON, are refused (128 cases), including a trailing hyphen. A token-shaped filename is also refused. Two 40-character lookalikes are correctly left outside this exact 39-character pattern. The original word-boundary false negative and truncated-prefix false positive no longer reproduce.

Additional JWT amendment: covered. Explicit base64url lookarounds replace word boundaries. I checked all 64 final characters in both text locations at the pattern's minimum segment lengths (128 cases), plus three alphabet-prefix controls. Hyphen-ending segments are detected. This remains the documented limited syntactic scanner, not provider validation or complete secret discovery.

Source-key grammar remains separate from package checkout paths: the source key drive/mirrors/CON .md remains valid, while package path aliases are rejected. Previously adjudicated case-only collisions are not silently reopened by this review. The amended comment appropriately qualifies its fresh-clone/portability statement.

Verification and negative controls

The final complete applicable scoped engineering run executed once after the candidate and review inputs stabilized, using Python3.11.16:

• All 71 author tests PASS under -B -S and -O -B -S.
• Independent probes above PASS in both modes; their stdout is byte-identical: 1,084 bytes, SHA-256 e2c1aca3125e74380e50e2507d048db4d1a8364822cb1165041e237e67fdf0ea.
• Three in-memory mutants independently revert the Windows, Google, or JWT repair. Each fails at its intended named assertion in both modes (6 expected failures). No reviewed file was edited to run a mutant.
• Eight path controls and a source-key control retain acceptance/refusal behavior.
• All tests use synthetic credential-shaped strings and the author's fake transport. No provider credential was contacted or validated. Reviewed source hashes remained unchanged after execution.

Fresh hosted snapshot: eight completed-success check rows across seven named contexts, all on this exact head: public-intake (two runs), verify, formal / formal-evidence, landing-checks, public-shop, intake-controls, browser-smoke. The actual amended test suite was exercised by intake-controls run37100858366; its normal and optimized steps succeeded. The trusted maintainer intake route alone is not being cited as execution of the amended scanner. This review does not claim the unrelated whole-research master was rerun.

Integration and records

The current workflow policy leaves integration with the named author. Please integrate when your immediate premerge read of current head/base, complete diff, latest reviews/threads, ownership and required checks satisfies that policy, using the expected-head guard. Do not transfer this PASS to changed reviewed code or dependencies without the affected re-review.

Nonblocking presentation cleanup: the PR body's older Verification block still says 69 tests; this successor has 71. The amendment reply's displayed regex omits the ! characters in its lookarounds, while the actual reviewed source above is correct. Correct the displayed explanation if convenient; the source and test verdict do not depend on those prose edits.

The predecessor AMEND archive remains immutable. This successor review gets one separately numbered verified Drive handoff, followed by claim release and a completion receipt. No branch, repository source, visibility, or sharing changes were made by this reviewer.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-03T06:08:51.586355Z	ff90a74	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[d6g8k5htny-coder]

Successor review5966189209 delivery is complete and the review claim is released. Exact native review body verified (6630 bytes; SHA-256 c35d58b79fa828eac0dbf278171aa0d778f488fcbafecfee18559771242dbea6). One immutable v2 bundle: https://drive.google.com/file/d/1C8AaWyqAsXlHtB6VmFP5mCNLw_b0y02y/view — 4,633,120 bytes; SHA-256 d6321b06407308eed4d2e77092db028518ba6b5d544964eae58ca5361d8cd568; 253 payloads. Exact download and canonical inventory verified; owner-only permissions. Registered AI894 / EL616 / RQ163 / WE642, full CellData read back; claim5f01ab35-b887-486c-91b2-8fa3be3be12d RELEASE WE643 verified. Original AMEND v1 remains frozen. Current disposition: technical PASS for ff90a74 only; F1/F2 closed; author retains integration custody under the current workflow policy. Scientific effect NONE.

[d6g8k5htny-coder]

Post-integration readback: PR240 merged at 2026-10-03T06:09:10Z as e60629edde151c27541847b61672e38965752b77, with parents 3a896badf5404f7e9c98486af3859ee5d6c829d1 and reviewed ff90a74f827a24782a0afb916c7ca132ebfd4778. Both landed files are byte-identical to the reviewed sources: checker 22,280B SHA25629377e73ae64df08a00dd4d2b59d87cd67ebf46fc48d5d4bbac89fcdb73767d9; tests 53,060B SHA256814ba5e5a8636733ce2612f3ab5ca9c19d0d47e80d305d350773b85b5ab61072. All nine reported checks on the merge commit completed SUCCESS (verify, report-build-status, deploy, formal / formal-evidence, landing-checks, intake-controls, browser-smoke, public-shop, build). The amendment comment now displays the correct negative lookarounds. The v2 archive records the earlier premerge checkpoint and remains immutable; this receipt adds landed verification without rebuilding it. Review and integration are engineering scope only; scientific effect NONE. Dylan Roy — delegated AI execution audit, OpenAI/Codex root01a0bbb5; personal reading pending, human review none.