# Site and repository reliability review — 2026-10-02 This implementation follows the authorized repository and website improvement pass. The baseline is public main commit `1badb9884253e4a1109e64b8fa27269c88685758`. The commit that publishes this report, its bounded control tail, and the Actions deployment receipt identify the final candidate; a source report cannot predict its own future Git SHA. ## Acceptance criteria and failure boundaries - Preserve all 219 existing work keys and reserve retired keys independently of editable titles and years. Reject duplicate, unsafe, missing, retired, or orphaned identities. Validate intake before source mutation and preserve concurrent edits when a failed update needs rollback. - Reclaim Pages headroom through a narrow projection policy while retaining all paper PDFs and manuscript files. Legacy QA captures remain in Git custody with manifest fallbacks; nested, invalid-date, and non-image report paths stay hosted. - Share progressive search loading, retain full search text for every record family, and recover from mixed deployment revisions without silently reporting a successful empty result. - Make mobile navigation usable at 320px without JavaScript, contain and restore modal focus, and report actual clipboard outcomes with a usable link fallback. - Keep the offline shell small and runtime cache bounded. Protect good cached responses from HTTP errors, storage failures, update churn, and stalled or oversized network bodies. Exercise these boundaries in real Chromium. - Fail required hosted browser/Lighthouse checks when tools, complete reports, or qualifying category scores are unavailable. Preserve diagnostics as Actions artifacts instead of treating missing execution as a pass. - Bind deployed manifest bytes to the committed candidate, check every work page and critical shared asset by SHA-256, check every archived PDF's HEAD contract, and check three deterministic PDF samples by SHA-256. Preserve the receipt outside the source tree with bounded retries and complete-response deadlines. - Invalidate cached generation when an owning script, shared source library, or declared imported orchestrator changes. Missing source inputs require a rebuild; always-run steps retain their explicit fixed-point behavior. - Regenerate source-dependent outputs in the declared order, confirm byte stability, run the repository suite and validation, and obtain fresh independent review of shared infrastructure changes before publication. ## Reviewed source changes [Bibliography metadata decisions](bibliography_metadata_review_2026-10-02.json) bind three full titles to archived title pages and the Active Blockference publication year to Crossref. Existing citation keys remain unchanged, including the historically named 2022 key for the chapter published in 2023. The Discovery Engine abstract is corrected against its archived source instead of perpetuating an unsupported free-energy synopsis. Human-facing search summaries and abstract search text also use the reviewed inert-prose normalization; raw enrichment and archived source bytes remain intact, including literal mathematical notation. [Repository catalog decisions](repository_catalog_review_2026-10-02.json) record five canonical URL migrations with matching immutable GitHub repository IDs and the pinned README source for OmniLatticeTextbook. The catalog description preserves the project's no-DOI status and does not adopt unsupported scientific-validation claims from repository prose. The canonical Scholar profile was observed directly in an authenticated browser on 2026-10-02: 823 citations, h-index 14, and i10-index 17; the corresponding since-2021 values are 594, 14, and 15. The snapshot and receipt bind the exact snapshot bytes and observation method. Independent review verified the binding against a sanitized observation excerpt; that reviewer did not independently operate the browser. Public artifacts exclude account identity and screenshots. ## Verification and provenance Local acceptance on 2026-10-02: | Check | Observed result | | --- | --- | | Initial forced ordered regeneration | Two passes; 100 local surfaces run, none skipped | | Final source follow-up regeneration | Two ordered passes; 69 surfaces run, 31 skipped with unchanged complete inputs | | `DOCXOLOGY_REQUIRE_BROWSER_QA=1 uv run python3 -m pytest code/tests -q` | 1,204 passed; zero failures, errors, or skips; 346.471 seconds | | `uv run python3 code/orchestrators/validate_repo.py` | Repository validation completed | | `uv run --group lint ruff check code` | Passed | | Workflow lint and `git diff --check` | Passed | | Static accessibility | 2,535 of 2,535 checked pages passed | | Artifact budget | 834.35 MiB against the 890 MiB gate; Pages projection 835.5 MiB | | Fresh external links | 853 observed: 699 OK, 146 rate-limited/bot-protected, one connection failure, three timeouts, four transient responses, zero HTTP 404s | | Independent cache review | 139 focused cases passed; no missing imported helpers across 31 cached stages | | Independent shared writer review | 70 focused cases passed; full-page stamp reuse and descriptor-safe, complete-map reads verified | The full test run included real Chromium lifecycle, progressive search, mobile/no-JavaScript navigation, focus, clipboard, first-paint geometry, and all eight real Lighthouse 13.4.1 reports. Thirty-two runtime/source hashes remained unchanged during the run. An earlier local Lighthouse process failed with a Puppeteer connection closure on Videos; its unchanged retry succeeded. That failure remains recorded separately; the final full run above passed cleanly without retries or runtime errors. All eight pages passed their existing floors with one Lighthouse run each: | Page | Performance | Accessibility | SEO | CLS | | --- | ---: | ---: | ---: | ---: | | Homepage | 75 | 100 | 100 | 0 | | Publications | 69 | 100 | 100 | 0.001407 | | Art | 74 | 100 | 100 | 0 | | Videos | 88 | 93 | 100 | 0.021526 | | Search | 72 | 100 | 100 | 0.110704 | | 404 | 99 | 100 | 100 | 0 | | Representative work | 97 | 100 | 100 | 0 | | Representative video detail | 95 | 100 | 100 | 0 | These scores do not establish the aspirational 85/95/95 targets on every page. Search's remaining initialization shift comes from wrapped type filters appearing after the core index loads; the footer shift was repaired. This measured follow-up remains under DOC-009. An extra ordered pass and clean post-landing validation are final publication gates. Hosted jobs and live acceptance belong to their candidate-SHA Actions receipts; local tests alone do not establish successful deployment. The fresh external-link capture records its dirty worktree honestly and is an outbound observation, rather than clean candidate release evidence. A pre-publication hosted run caught an ambiguous clipboard-test selector after asynchronous search results added more headings. The regression tests now scope clicks and completion assertions to the intended Research heading while keeping the real search rendering, pending clipboard promise, query-preservation, and unavailable/denied clipboard checks. The failed run is retained in Actions; the replacement candidate must pass the complete hosted gate before publication. Independent source review checked the title pages, Crossref publication year, GitHub repository identities, pinned README hash, public-text privacy, all active work keys, retired reservations, and Scholar snapshot binding. Independent code review exercised shared search, service-worker, intake, and deployment-verifier failure cases; its findings were repaired before landing. Footer stamps record rendering provenance and retain their date only when the complete page is unchanged except for the stamp. Shared writers and checks now use the same rule, validate the complete output mapping before reads, and finish all descriptor-safe reuse reads before writes. They do not prove that a future publication commit contains the rendering. Exact deployment SHA and live bytes are separately bound in the retained technical acceptance receipt. That receipt does not claim all artifact files were downloaded, human visual review occurred, or the complete release attestation passed. ## Remaining source and access work The [active backlog](../TODO.md) retains recurring maintenance and concrete external dependencies. Licensed full text remains unavailable for ToComment, Paleolithic Rockstars, and Focused Attention Meditation. Digital Twins and Aligning AIO to SUMO author discrepancies require depositor reconciliation; registered creators are preserved. Contradictions inside originating manuscripts require source publication changes before archived summaries can adopt them. The available signed-in browser could not access the site's Search Console property. Owner-granted property access is required for sitemap and indexing follow-up; no DNS, ownership, or account settings were changed. Flickr-side enrichment remains an account editing task. A managed-profile deep security scan is a separate evidence requirement and is never inferred from ordinary tests or independent code review.