Review: harden AKC mirror from PR #40

[jackjin1997]

Reviews and preserves the contribution from #40. The original AKC commit remains
authored by @Sun-sunshine06; reviewer fixes are appended on this branch.

Detailed review: review-reports/PR-40-AKC.md

HF asset candidate: https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/97

Executed visual/task fixed point: ea6c2cbeb9a55f5c8f0a03a9a21130debef4d9ee

Included

• source-shaped responsive UI using 40 pinned local images
• reproducible immutable seed/archive and registry integration at port 40040
• validated account, selector, saved-breed, profile, and event-registration behavior
• 13 reviewed tasks with offline deterministic verifiers
• refreshed public visual, task, and engineering evidence

Executed

• 88 fixed-viewport source/mirror captures across 11 task-path pages; mirror 44/44 without horizontal overflow
• 13/13 post-repair real Playwright UI runs completed and passed deterministic grading
• 8/8 AKC app tests; 13/13 verifier methods covering 71 subtests
• 58/58 anonymous routes returned HTTP 200
• 41-site registry and full extracted-asset checks passed
• immutable HF archive re-downloaded and validated at revision dc253d3f17cb5d83ee00f9bbc2100a296126327d

Draft blockers

• replacement independent blind review is pending; the earlier result predates the material visual repair
• fresh full Docker build is not run: the host has 35 GiB free and the review procedure requires 50 GiB before starting

Nothing should be merged from this Draft. I will request maintainer review only after
both remaining gates pass. Do not merge either PR from the reviewer account.

Links back to original PR: #40

[jackjin1997]

Independent blind review — first pass (frozen packet)

Model: claude-opus-5 (session-declared; not independently attestable from inside the session)
Reviewed at: 2026-09-18T12:52:20Z
Packet manifest SHA256: d43cdb643cd3ddcd60391005e2d9ed0ab9e4922864250c93eb01baf73afdff5a — manifest hash matched the expected value; all 125 allowlisted files re-hashed and matched (0 mismatches, 0 missing).
Executed code SHA: 8a06cfae154ec100feb00b194c2a64389a6756dc
Asset revision: 4b3879d288dabff17257e04b3b27e43b8628f63d · Asset archive SHA256: a06876a6fb239a340f03134bc66c443ffbaa5a1df43e739eb914cb338bc3f699

Coverage: 13 of 13 tasks (AKC--0 … AKC--12), one frozen run each (candidate-001).
Result: 13 PASS / 0 FAIL.
Failures: none.

Method: each verdict rests only on the frozen task text, the non-answer-leaking rubric, the full trajectory and final answer, step screenshots, and a read-only initial.db vs after.db diff across all six business tables. Read-only tasks (AKC--0..7, AKC--12) showed no row added, removed or modified. Write tasks matched their rubric exactly: AKC--8 one event_registration row (user 1 alice_j / event 3 puppy-training-webinar / Scout / Canine Good Citizen); AKC--9 a single field change user.activity_level Moderate→Low on Alice with household, experience and all other data untouched; AKC--10 one saved_breed row (user 2 bob_c / breed 19 whippet); AKC--11 one new user row with all specified fields and a scrypt password hash rather than plaintext. No packet file was modified.

Unexecuted checks (explicitly not covered):

• Full Docker build / environment bring-up — packet records NOT_RUN (44 GiB free, 50 GiB required); the site was never re-run.
• Deterministic verifier source and results, hidden answer keys, prior reviews, reviewer conclusions and PR discussion — excluded from the packet and not read before this result was hashed.
• Source authenticity / asset provenance — SHAs above are recorded values, not independently re-derived upstream.
• UI fidelity against the real akc.org — not assessed.
• Model success rate, reproducibility across seeds or reruns, task-difficulty calibration — only one run per task was available.
• Screenshot-to-DOM binding — screenshots spot-checked for 3 of 13 runs; the rest were judged from trajectory text plus SQLite state.

Limitation: this is a blind judgement of whether each frozen execution completed its task, and nothing more. It is not a certification of source authenticity, UI fidelity, verifier implementation correctness, a full Docker build, or model success rate.

Structured result: cases/WHR-040-akc-r2/independent-review/result.json, SHA256 565d9bd94e0805183b0102a83ff8c26cbb61c1b769349251fbae582da80a1d36 (frozen before any non-allowlisted material was read).

[jackjin1997]

Visual-repair update: the Owner rejected the previous home-page fidelity. The branch now carries a source-shaped repair at ea6c2cb, with 88 fixed-viewport source/mirror captures across 11 task-path pages (44/44 mirror captures without horizontal overflow), 13/13 refreshed real UI runs passing deterministic grading, and HF asset revision dc253d3f17cb5d83ee00f9bbc2100a296126327d. The earlier blind result is historical only because it predates these material UI changes. A replacement 125-file blind packet is frozen and pending Owner execution. The PR remains Draft; the fresh full Docker build is also still blocked by the 50 GiB free-space precondition.