Rebuild the site on Astro, replacing al-folio - #6

Same content, same URLs, different machinery.

ardoco.de ran on a pre-v1, vendored-theme al-folio fork. Upstream has since moved to a
gem-based v1, so the fork sat on a branch nobody maintains any more: 27 gems, 14 workflows and 20 MB
of build-time-downloaded vendor libraries, to serve about 25 MB of actual content. fuchss.org hit
the same wall last week and wrote it up; this
ports that approach to ARDoCo, with its own design.

Data errors now stop the build

Content lives in six Zod-validated collections — publications (16, parsed from papers.bib at
build time), conferences (16: 11 paper pages + 5 redirect stubs), approaches (8), people (6),
authors, venues.

The joins that Jekyll resolved as string lookups are Astro reference()s now, so a conference
naming an approach that does not exist, a BibTeX abbr with no venue entry, or an
html = {/c/<slug>} pointing at a missing page fails the build instead of rendering blank.

That surfaced two latent bugs immediately:

• abbr = {SE} had no venues.yml entry, so those badges rendered grey and unlabelled.
• The children[] array in _pages/conferences.md silently drove both the navbar order and the
  front-page publication order, so the two could drift. Each conference entry now carries its own
  navOrder.

URLs are unchanged

Checked against the live gh-pages branch during the migration; every published path still
resolves.

• build.format: 'preserve' emits both shapes from one config: /approaches/<slug>/ as a directory
  and /c/<slug> as a flat .html, exactly as Jekyll did.
• The five legacy stubs (se24, se25, se26-exarch, se26-lissa, taas26) still redirect to the
  papers that superseded them.
• All 32 PDFs and images are served byte-for-byte from public/ and pinned by SHA-256.
• public/CNAME is now committed. It previously survived only because Jekyll's keep_files
  preserved it on the branch; a full-replace deploy of a new tree would have dropped the custom
  domain and left the site reachable only at ardoco.github.io.

/_pages/conferences/ — an accidental Jekyll address, published because that page had no
permalink: — is retired in favour of a real /conferences/ index.

Bibliography fixes

• Four GI "Software Engineering" papers were typed @article with the publishing society in a
  journal field. They are proceedings: retyped @inproceedings, bogus field dropped, LNI series
  named on the two carrying a P-nnn volume.
• Non-breaking spaces hidden inside two titles. Invisible in an editor, but they broke word
  wrapping, were copied into anyone's .bib via the BibTeX button, and stopped the publications
  filter matching a typed space.
• A month = {04}, two raw ü where the rest of the file uses LaTeX escapes, and a duplicate ISSN.

Crossref still reports print-vs-electronic ISSN differences and shortened venue names — both are
correct as they stand, and the GI/ACM 2026 DOIs are simply not Crossref-registered.

Also new

• A real /conferences/ index; the 11 paper pages were dropdown-only before.
• Highwire citation_* tags and ScholarlyArticle JSON-LD on every paper page — the old site
  emitted neither (serve_og_meta and serve_schema_org were both false).
• Email addresses appear nowhere in the served bytes, raw or entity-encoded, while still reading
  as hey (at) kit (dot) edu without JS. The old jekyll-email-protect output survived one
  unescape call.
• Dark only. No theme toggle; the audit fails if one reappears.

Review

npm install
npm run dev      # http://localhost:4321
npm run build
npm run verify   # 43 checks

npm run verify asserts: the 32 assets are byte-identical, CNAME and .nojekyll are emitted, all
568 internal links and 191 fragments resolve, WCAG contrast holds, all 50 images carry intrinsic
dimensions, no text runs into a link, no address leaks, www.youtube.com (the two screencasts) is
still the only third-party origin, and the two NEVER CHANGE THIS LINE B/E comments that
mcse.kastel.kit.edu scrapes are present verbatim with their three paragraphs intact.

Worth a look by hand: /c/icsa25 (the richest page), /c/taas26 (should bounce to icsa25),
/people/#dominik-fuchß (anchor reached from a publication author link), and /approaches/lissa/
(reverse-joined publications).

Notes

• Merging deploys. The workflow runs prettier --check, astro check, build and verify,
  then publishes dist/ to gh-pages.
• astro check reports 2 hints in src/pages/c/[slug].astro, claiming two used imports are unused.
  That is a checker limitation with the early return in Astro frontmatter; exit code is 0 and the
  built stubs prove the branch runs. Documented in the file.
• The screencasts embed YouTube directly, as before, so those two pages load a third-party origin.
  It is an explicit allow-list entry in the audit, so a second one cannot appear unnoticed.
• The "~30 s" old build time is an estimate — the Jekyll build was never run locally.

Follow-ups (not in this PR)

• 11.7 MB of the repo is four .pptx decks that duplicate their own PDFs.
• icsa26-archguard.svg (1.1 MB) and aire25-aire.svg (711 KB) are oversized; optimising them
  changes bytes at pinned URLs, so it is a deliberate separate decision.
• AIRE 2025 and REFSQ 2025 render without a venue badge — those bib entries carry no abbr, which
  was true before this PR too.