v1.4.0

Added

• astro-seo.md (EN/JP) — Astro-specific SEO layer: client:* directives mapped to INP/LCP
  costs, server:defer server islands, canonical and Open Graph generation from
  Astro.site + Astro.url, JSON-LD generated from Content Collections schemas (with
  </script> escaping), @astrojs/sitemap configuration including multilingual sitemaps with
  hreflang, view transitions (native vs. <ClientRouter />) and their analytics pitfalls, and
  version migration notes for the SEO-affecting breaking changes in Astro 7 (Rust compiler HTML
  validation, compressHTML: 'jsx' collapsing whitespace in indexable text, the Sätteri Markdown
  pipeline) and Astro 6
• edge-seo.md (EN/JP) — Cloudflare Workers/Pages SEO layer: _redirects design and limits,
  _headers with X-Robots-Tag and cache strategy, the static-asset vs. Worker-code blind spot,
  dynamic sitemap.xml generation from D1/KV with edge caching, crawler verification by published
  IP ranges and Web Bot Auth, HTMLRewriter metadata rewriting and JSON-LD injection, and the
  relationship between edge caching and crawl budget
• last_verified frontmatter on every skill markdown file, enforced by CI
• docs/research-notes.md — the verification log behind every factual claim, with source URLs
• .github/workflows/freshness-reminder.yml — opens a freshness-check issue on the 1st of each
  month with the list of sources to re-verify
• .github/workflows/release.yml — publishes a GitHub Release on v* tags, using the matching
  CHANGELOG section as the release notes
• Versioning policy section in both READMEs, and Astro / Cloudflare Workers/Pages in Supported
  Frameworks

Changed

• Skill descriptions now include Astro and Cloudflare Workers/Pages invocation triggers, and state
  which of the two skills to load: English input selects seo-mastery, Japanese input selects
  seo-mastery-jp. Previously the two descriptions were near-identical, so either could fire
• SKILL.md (EN/JP) gained an "Astro and Edge (Cloudflare) Specifics" section, registered the two
  new reference files, and now states that topic files and platform files must be read together —
  sitemaps, hreflang and crawl budget each span more than one file
• audit-workflow.md (EN/JP) points at the platform files, since several audit phases surface
  findings whose cause is platform-specific
• README positioning no longer claims freshness is "guaranteed": what is guaranteed is the process
  and the date stamp. Supported Frameworks now links each entry to the file that backs it, and drops
  the WordPress entry, which had no supporting content anywhere in the repo
• CONTRIBUTING.md's Japanese section is a full translation rather than a summary; it previously
  omitted the repository layout, the one-language-PR rule, the freshness workflow, and all four
  content guidelines
• marketplace.json gained astro, cloudflare-workers and edge-seo keywords
• CHANGELOG restructured to Keep a Changelog format; historical entries re-dated from git history
  (the v1.0.0 entry previously read "2025-01"; the initial commits are dated 2025-12-29/30)
• scripts/validate.py now verifies last_verified on every skill markdown file (present, ISO
  format, not in the future, identical between EN and JP) and ignores # lines inside fenced code
  blocks when comparing document structure

Fixed

• AI search eligibility was incomplete. ai-search.md and SKILL.md stated that indexability
  plus snippet eligibility were the only requirements for AI Overviews / AI Mode. Google's
  generative AI optimization guide (2026-05-15) adds a third: the site must be included in Search
  generative AI features in Search Console. Documented the Search generative AI control
  (Settings → Search generative AI, rolled out 2026-06-03), its three options, and the fact that
  Google states it "isn't used as a ranking or inclusion signal affecting other parts of Search" —
  making it the correct tool for AI-only exclusion, unlike nosnippet
• ai-search.md now cites the 2026 generative AI optimization guide, including its verbatim
  mythbusting on llms.txt, content chunking, AI-specific writing, and AI-specific schema
• Search generative AI performance report described with its actual limits: no historical backfill,
  impressions only. The 2026-05-18 data start is attributed as a secondary-source figure, since
  Google's own documentation does not state one
• technical-seo.md sitemap best practices: noted that Google ignores <priority> and
  <changefreq>, that lastmod is only used when consistently accurate, and added the sitemap
  index limits
• technical-seo.md crawl budget: removed "parameter handling (Search Console)" — the URL
  Parameters tool was retired in 2022
• technical-seo.md gained the robots.txt limits Google actually enforces (500 KiB, 24-hour cache,
  crawl-delay and noindex unsupported, longest-match precedence)
• technical-seo.md Nuxt SSR example replaced Nuxt 2 asyncData + $axios with Nuxt 3+
  useAsyncData / $fetch
• core-web-vitals.md Nuxt example replaced render.http2.push (HTTP/2 Server Push, removed from
  Chrome) with a Nuxt 3+ @nuxt/image config and explicit LCP preload
• core-web-vitals.md corrected the scheduler.yield() support note: Chromium and Firefox, not
  Safari, not Baseline
• audit-workflow.md robots.txt check no longer asks whether Crawl-delay is set too high; it now
  flags crawl-delay / noindex lines as ineffective for Google
• audit-workflow.md AI readiness phase now checks the Search Console generative AI control

Found by review of this branch before release:

• Article headline has no 110-character limit. Google removed it from the documentation in
  January 2023; the current wording only warns that long titles may be truncated. The claim was
  presented as a Google requirement in structured-data.md (since v1.0.0) and astro-seo.md
• _redirects placeholders do match dots inside the path. The delimiter is / in the path and
  . or / only in the host, so /:slug also matches report.pdf — the previous wording made
  such rules look safe for file URLs
• _headers rules accumulate rather than override, joining repeated header names with a comma.
  The per-crawler X-Robots-Tag example would therefore have shipped one combined header that
  Google does not document parsing; removed, with the collision rule documented. The /*.html
  cache rule was dropped too — under auto-trailing-slash it matches nothing
• Middleware-style Workers need run_worker_first. Cloudflare serves a matching static asset
  before invoking the Worker, so the HTMLRewriter and query-canonicalization examples could never
  have executed under the config shown
• Supplied the ipInCidr implementation the crawler check called but never defined; made the range
  lookup fail open rather than 500 on a fetch error; stopped an empty KV value poisoning the cache
• Dynamic sitemap generation no longer 500s on a NULL or non-ISO updated_at, strips
  XML-illegal control characters, and cannot emit a zero-entry sitemap index
• HTMLRewriter handlers now catch their own exceptions (a throwing handler errors the response
  body), guard missing values, and scope the selector to head > title so inline SVG titles are
  left alone
• Removed the unsourced claim that 410 is processed faster than 404; attributed the generative AI
  report's 2026-05-18 data start as a secondary-source figure and removed it from SKILL.md;
  replaced "expanding from July 2026" and "takes effect within 1–2 days" with Google's own wording
• freshness-reminder.yml declared only issues: write, which sets every other scope to none —
  actions/checkout would have failed on the very first run
• validate.py crashed with a traceback on a syntactically valid but impossible date, could fail in
  CI for a file stamped "today" in JST, mispaired indented code fences, and reported one malformed
  file three times. It also now checks that every declared version agrees — v1.2.1 and v1.2.2
  shipped with both manifests still reading 1.2.0
• release.yml's empty-section guard could never fire (the extracted section always contains a
  newline), the notes regex now terminates on end-of-file rather than relying on the link-definition
  block, and a re-run updates the release instead of failing on "already exists"

Found by a second review pass on the same branch:

• The dynamic sitemap published 404s for hierarchical slugs. locFor() ran
  encodeURIComponent() over the whole slug, so 2026/my-post became 2026%2Fmy-post — a
  different URL, submitted to Google as canonical. It now encodes each path segment
• astro-seo.md's blog route overrode the canonical with a hardcoded trailing slash. The
  getStaticPaths sample built `/blog/${post.id}/` and passed it as canonicalUrl, so under
  trailingSlash: 'never' every post declared a canonical and og:url that did not match the URL
  it was served at — the two-URLs-per-page split the next section warns about. The sample now lets
  BaseLayout derive the canonical from Astro.url.pathname
• The tracking-parameter 301 rewrote parameters it was not meant to touch. Mutating
  url.searchParams re-serialises the whole query, turning ?q=a%20b into ?q=a+b; an origin that
  normalises it back would bounce the request between the two rules. It now rebuilds the query from
  the raw pairs
• The sitemap Worker called cache.put() without checking the method, so a HEAD from any crawler or
  uptime monitor rejected inside waitUntil; it also ran a cache lookup before routing, on every
  static asset request, and spent a second COUNT(*) query per page purely to bound the page number
• astro-seo.md described a missing site as silently producing broken canonicals. new URL(path, undefined) throws, so the build fails — and only one of the three components that dereference
  Astro.site had an explicit guard, which made the other two look like the silent case
• The @astrojs/sitemap serialize example demonstrated priority, which the paragraph below it
  says Google ignores, and matched only the blog index rather than the posts. It now sets lastmod,
  the one field the file says is worth getting right
• validate.py rejected valid frontmatter: block scalars, sequences, nested mappings and multi-line
  plain scalars all errored as "not key: value", so adding any standard multi-line field would
  have failed CI. It also missed a value ending in a colon — the same scanner error the plain
  scalar check exists to catch — false-positived on values starting with -, ? or :, which YAML
  accepts, and left \uXXXX escapes unexpanded while measuring description length against them
• release.yml re-implemented frontmatter parsing in sed, which disagreed with validate.py over
  quoted values: version: "1.4.0" passed CI and would then have failed the release, after the tag
  was already immutable. The tag check is now validate.py --expect-version
• freshness_table.py walked the skills tree recursively while validate.py globbed one level, so
  a file in a subdirectory would be demanded in the reminder issue but never enforced by CI. Both
  now walk through the same function, the table exits non-zero on malformed frontmatter instead of
  rendering it as "missing", and CI runs the script so a rename on its import cannot go unnoticed
  until the monthly workflow fires