Metadata recall + content quality improvements (defuddle-inspired)

[dazld]

Summary

Five focused improvements to metadata extraction and content cleaning, drawn from a comparison with defuddle. The bigger architectural shifts (extractor registry, scoring-based main-content discovery) are deferred.

• :image as a first-class metadata field — sourced from JSON-LD image (string or {url} object), then og:image, then twitter:image. Fixes the markdown-negotiation follow-up where upstream image: frontmatter mismapped onto :icon.
• Placeholder filter for metadata strings — rejects unresolved template literals ({author.fullName}), anchor placeholders (#author.name), and decorative strings without letters/digits at the leaves (safe-attr, safe-text, get-json-ld-value, get-json-ld-image).
• og:site_name word-count guard — sites occasionally put the article title in og:site_name. Reject anything over 6 words and fall through.
• JSON-LD @graph walking + multi-script + entity decode — read all <script type=application/ld+json> tags (not just the first), strip /* */ and // comments, flatten @graph arrays, prefer Article-typed objects, recursively decode HTML entities.
• Semantic-div standardization — convert div[role=paragraph] → <p>, div[role=list] → <ul>, div[role=listitem] → <li> early in clean-document so React/Next.js sites with role-based markup produce real paragraph/list output.

Test plan

• clojure -M:test — 99 tests / 255 assertions, 0 failures
• New tests cover each commit (image extraction from og/twitter/JSON-LD/object, placeholder rejection, sitename guard, @graph walking, multi-script preference, entity decoding, JSON comment stripping, role-div conversion to <p>/<ul>/<li>)
• clj-kondo — no new warnings
• Manual smoke tests against the GraalVM native binary:
  • https://developers.cloudflare.com/docs-for-agents/index.md — image: populated from upstream YAML, icon: correctly empty
  • https://www.bbc.com/future/article/... — :image distinct from :icon (article hero vs favicon)
  • https://www.wired.com/story/... — same; clean title/author/description from JSON-LD
  • https://www.dbreunig.com/... — single-script JSON-LD path, no regressions
• Manual smoke test on a site known to use <div role=paragraph> (worth doing before merging — none of the smoke URLs above exercise this code path, only the unit test)

Deferred

Captured for follow-up:

• Schema.org articleBody fallback as content source
• Tiered retry strategy on low word counts
• H1-adjacent date/byline scan
• Extractor registry pattern (architectural — separate effort)