v0.7.0

Added

• tests/test_search_pages.rb covers the generator: both pages when search is on, neither when it is off or unconfigured, and no duplicate when the site provides its own.
• The README carries a badge showing the version published on RubyGems, linking to the gem page.
• scripts/verify_gem_package.rb checks that a built gem contains every bundle its manifest references; the release workflow builds the bundles and runs it before publishing.
• tests/test_gem_package.rb covers what a site using the gem needs: the bundles are packaged, the plugins' gem dependencies are declared, and requiring the theme registers its Liquid tags.
• The search index test checks that tags are whole tags rather than only that the field is an array.
• The integration suite runs axe-core over six pages in both light and dark mode, so contrast, missing accessible names and misplaced ARIA cannot regress unnoticed.
• A social card image (assets/img/social-card.png), so the Open Graph and Twitter image tags no longer point at a missing file.
• The release workflow now does a release in one run plus one pull request: "Run workflow" with a version bumps develop and opens the PR into main; merging it tags main, publishes the GitHub release and starts the gem publish.

Changed

• MathJax is only loaded on pages that contain math (theme_options.math.render_on_load: auto, the new default of the demo site) and Prism only on pages with a code block (theme_options.syntax_highlighting.load: auto); pages can still opt in or out with math: and syntax_highlighting: front matter, and a page with math: false is also left alone by the math preprocessor. Pages without either skip about 400 KB of CDN scripts and stylesheets.
• The IBM Plex web fonts fall back to local fonts scaled to Plex's metrics, so the swap once the web font arrives no longer moves the layout, and the Google Fonts stylesheet no longer blocks rendering.
• The home hero image is preloaded and small screens get a 640 px variant (hero_image_small for pages that set their own hero_image).
• The Lighthouse workflow inlines critical CSS before building, as the deploy does, so it measures the published configuration.
• CI jobs install only what they use: the Jekyll test job no longer installs libvips (the image plugin uses MiniMagick, which the runners already provide) or keeps redundant pip and _site caches, and the accessibility, Lighthouse and deploy workflows no longer set up Python, which only the Rake verification task needs.
• The hero background is served as a 46 KB WebP instead of a 1.27 MB PNG.
• Footer text and links, and the skip link in dark mode, meet the 4.5:1 contrast ratio; the blog listing uses second-level headings for its cards.
• The Tests workflow runs for pull requests into main as well as develop; the duplicate theme-stability workflow is gone.
• The configuration guide documents where settings actually live (_config.yml plus _data/config/author.yml).
• The gem publish can authenticate with RubyGems trusted publishing (OpenID Connect) instead of a stored API key; the RUBYGEMS_TRUSTED_PUBLISHING repository variable selects it.

Fixed

• Sites installing the gem had the search interface and its JavaScript but no search: the page that renders it and the one that builds its index are pages, which a theme gem cannot ship. Both are generated now for any site with features.search enabled, and a site that defines either path keeps its own.
• The published gem could not be used. Jekyll reads _plugins/ for a site but not for a theme gem, and lib/datalog-theme.rb did not load them, so the tags the layouts use were never registered and every consumer site failed to build with Unknown tag 't'. Naming the theme under plugins: now loads them.
• The gem shipped _data/js_manifest.json, which points every page at the browser bundles, without the bundles themselves: assets/js/dist/ is build output that git does not track, and the gemspec selected files with git ls-files. Sites using the theme loaded no JavaScript at all.
• The gemspec did not declare loofah, which the notebook plugin requires, so loading the theme raised cannot load such file -- loofah.
• The home layout sorted site.portfolio, which is nil unless a site declares that collection, and sorting nil stops the build. The section is skipped when the collection is absent.
• The search index shipped every tag as a list of single characters, so the tag filter on the search page offered 34 buttons reading a, b, [, " and so on instead of the 48 real tags. A split: '' in the index template turned each tag array into its string form before splitting it.
• A tag consisting only of whitespace produced a filter button with no accessible name at all, which fails WCAG 4.1.2.
• The fallback author avatar carried an aria-label on a plain div, where ARIA prohibits it, so screen readers announced nothing for it. It is now an image role, matching the photo it stands in for.
• Text set in the teal accent failed WCAG AA everywhere it appeared, reaching only 3.0:1 to 3.5:1 on post badges, skill tags, search highlights and Prism keywords. Teal text now uses a darker tone; the original teal stays for fills and borders.
• Dark mode had several unreadable components, none of which any check covered: the citation tools kept light backgrounds under light text, leaving a heading white on white at 1.09:1, and difficulty badges kept their light-mode text colour on a near-black pill at 1.93:1.
• Buttons marked btn--ghost were never styled, so they fell back to the browser's own button chrome and could not follow the theme.
• The Playwright and coverage reports were copied into the built site and published with it. Both are excluded now, and the Playwright report is also ignored by git.
• datalog check reported every dependency as missing on Windows, including bundler, which it treats as a critical failure. The probe called command -v, a POSIX shell builtin with no executable behind it; it now searches PATH itself, honouring PATHEXT.
• The build and test scripts run on Windows. npm run test:integration spawned the Playwright .cmd launcher, which Node refuses with EINVAL; npm run build:critical spawned bundle, which is a .bat there and failed with ENOENT; bundle exec rake ci:verify invoked python3, which on Windows is a Microsoft Store stub rather than the interpreter; and the CLI tests ran the binstub through its shebang. Node dependencies now run under the current Node binary instead of their launcher shims, and the remaining launchers go through the command interpreter with arguments quoted, so paths containing spaces survive.
• The site navigation no longer renders expanded and then animates shut on small screens once the script runs, which moved the whole page by about 340 px and put every page's cumulative layout shift near 0.3.

Removed

• Percy and its visual suite (tests/visual/): Percy never ran without a token and carried the last open npm advisory and about 150 packages, and the suite failed 34 of its 58 specs on CDN waits and strict locators, locally and in CI. The integration specs under tests/integration/ remain the browser checks and gate every deploy. npm audit is clean.
• The jekyll-jupyter-notebook gem and the Jupyter toolchain. Notebook pages are rendered by the theme; only nbformat remains, for the notebook validation script.
• _data/config/site.yml, theme.yml and features.yml, which nothing read.

v0.6.1

Fixed

• An overlay hero on the home page produced an empty id (and an "Empty slug generated" warning) because / slugifies to nothing; it now uses post-hero-home.

v0.6.0

Added

• tocify Liquid filter (_plugins/toc_filter.rb): the table-of-contents includes called content | tocify but no plugin defined it, so the aside rendered the whole article. The filter builds nested toc-list lists from the rendered headings within the requested level range.
• show_author: false in front matter hides the byline in the article meta, for pages that are site chrome rather than articles.
• The footer's "Explore" column reads footer: from _data/navigation.yml when present, falling back to the demo sections.

Fixed

• Prism was blocked on every page by a stale integrity hash for prism.min.js in _data/cdn-integrity.yml; the hashes are regenerated.
• The <head> emitted every SEO tag twice (title, description, Open Graph, Twitter, canonical, JSON-LD) because {% seo %} duplicated the theme's own markup; the tag is gone.
• The Content Security Policy blocked MathJax's accessibility data (connect-src now allows jsDelivr) and logged an error for a report-uri directive that <meta> policies cannot carry.
• The home-page hero set its background through a style attribute the policy blocks; it now uses a nonced style element.
• The header's search row overflowed the viewport on phones; it wraps below the brand now.
• <body data-theme> stays in sync with the effective theme when either theme control is used.
• The visualizations gallery's Plotly block gained the status, meta and toolbar markup the script and the docs expect.
• The <head> now writes the <title> itself for every page (Page | Site, or the site title on the home page); it used to rely on the removed SEO tag when no seo_title was set.
• Google Analytics was loaded with an empty measurement id because google_analytics: "" is truthy in Liquid; the tag is emitted only when an id is set.
• Google Fonts CSS and gtag.js are exempt from subresource integrity (both are served per-browser); the generator also ignores scratch directories, and the validator skips the notebook iframe documents nbconvert writes.
• Internal links written as root-absolute paths (the 404 page, the getting-started and tools pages, two posts, and the portfolio, post, package and project layouts) now go through relative_url, so the site works under a base path. A browserconfig.xml reference to a file that does not ship is gone.
• The deploy workflow builds the published artifact with the Pages base path, and runs the integration suite without the Percy-only @visual specs. npm run test:integration excludes those specs too, and the runner can start its static server on Windows.
• The archive layout grouped posts with where_exp expressions containing filters, which Jekyll's Liquid cannot parse; posts are now grouped by year and month with plain loops. Posts without a difficulty no longer trigger "Empty slug generated" warnings.
• The page layout printed the title and subtitle a second time above the article scaffold, which already renders them.
• The math preprocessor hooked both documents and posts, so every post was processed twice and display math came out nested. The inline pattern also matched the $ of a $$ block; it now requires single delimiters on both sides.
• {% t %} with a quoted key leaked the closing quote into the first option name, so {% t 'post.reading_time' minutes=... %} printed a literal {{minutes}}.
• Overlay post heroes set the background through an inline style attribute, which the Content Security Policy blocks. The image is now applied by a nonce'd <style> element scoped to the hero's id.
• Card thumbnails on card grids were stretched by the height attribute the image optimizer adds and by the grid stretching card bodies. Cards now keep their intrinsic thumbnail ratio and align their content to the top.
• The skip link peeked into the viewport before it received focus; it is now moved fully out of view until focused.