Performance

Performance

The goal is a fast page on a mid-range phone on a slow connection, reached by sending very little: no framework, no third-party request from the site's own assets, one small stylesheet. (The host injects its own analytics beacon at the edge; the CSP and the Privacy page name it.) The numbers are enforced by two scripts on every pull request that changes the site, not by a one-time audit. A documentation-only pull request skips them (docs/ci-parallel.md). The source of truth is docs/performance-gates.md; this page is the map.

What keeps the pages light

• No JavaScript by default. Three scripts ship, each under 1 KB gzip and each only on its own pages (Support, Search, and posts and journal entries). Everything else is HTML and CSS: the phone menu is a native popover, accordions are native <details>.
• No third-party requests from the site's own pages. The Content Security Policy forbids them, apart from the host-injected analytics beacon, and the build fails on any other. Zero external fonts, scripts, styles, images or embeds.
• One stylesheet per page, concatenated, minified and fingerprinted by Hugo Pipes. Page-specific CSS lives in assets/css/pages/, and the penguin ghost image used only on Home and 404 is its own file appended on those two pages, so a post is not charged for it.
• Two fonts, Jost for body text and Orbitron for headings, both served from the site and preloaded in the head.
• Images are resized by the build. The render hook and the headline-image partial write srcset, width and height and lazy loading (the hero is eager and fetchpriority="high"), and the headline image is served only as resized WebP copies. See Images and alt text and Headline images.
• Search is part of the page, not a separate index request.

The budget, in short

Enforced by scripts/check-budget.py (static, no browser). The limits:

Item	Limit
Scripts	Only the pinned ones, 1 KB gzip each
Stylesheet	One per page, 20 KB gzip
Fonts	At most 2 files per page, 30 KB each
Page HTML	15 KB gzip
Any image a page loads	120 KB
Any image file in the build	300 KB
A page, cold transfer	150 KB
A post without its body images	80 KB

The full table, today's measurements and the fixed weight every post pays are in docs/performance-gates.md. See also Build checks and budgets.

Measured in a browser

scripts/browser-gate.py serves the built folder on 127.0.0.1 with the real _headers, starts headless Chrome and loads each template at several widths. For performance it fails when:

Rule	Limit
Largest Contentful Paint (cold)	over 2000 ms
Cumulative Layout Shift	over 0.05
A failed or blocked request, or any request that leaves the local server	any
A console error or a CSP violation	any
Horizontal overflow	any, at six viewports including 320 CSS px and 200 percent zoom
A visible raster image drawn wider than the file the browser chose	any

LCP and CLS are measured under a throttled profile (1.6 Mbps down, 150 ms round trip, 4x CPU slowdown), the same shape as Lighthouse's mobile preset. The gate waits for the page and its fonts to load and settle before it reads the layout shift. Fonts use font-display: optional with a preload so a late font never reflows the text. docs/performance-gates.md lists what the gate does not prove: these are lab numbers on a fixed profile, not field data, and the Chrome version is not pinned.

Lighthouse and PageSpeed

Lighthouse itself is not a CI check, but a pass was done and recorded in docs/performance.md (issue #232, 2026-11-08, Lighthouse 13.6.0). On a local production build, desktop and mobile scored 100 for accessibility, best practices and SEO on every page except the 404 page, whose SEO score is 69 on purpose because it is noindex. Performance was 96 to 100; mobile differences of a few points between runs were noise. The pass fixed the About mascot, which was drawn wider than its file, by adding a srcset, and added an image rule to the browser gate that fails any visible raster image drawn wider than the file the browser chose. The public staging site scores SEO 69 by design (noindex) and Best Practices 92 because of a blocked analytics beacon request; those are owner decisions listed in docs/performance.md.

Tips for authors

• Resize images before you commit them and prefer WebP.
• Do not embed. Link.
• Check a long post against the 80 KB post limit with scripts/check-budget.py on a build.