Colophon

Typefaces

Four faces, each with a job. These samples use the same font files as the rest of the site; the names link to their specimens.

next/font supplies the screen fonts. The resume uses static Inter and Source Code Pro files from Fontsource because Chromium was drawing the variable-font glyphs as paths in the PDF. The measurements and fix are in Every Letter in This PDF Is a Drawing.

From File to Page

Next.js and React provide the pages. The article pipeline (opens in a new tab) runs on the server:

• Figures and code: captions, heading links, syntax highlighting, and diagrams are prepared with the text.
• Interactive explanations: the camera simulation and byte diff let you drive the thing the post is explaining.
• Reading tools: image inspection, copy buttons, and marginal notes run in the browser.

The Lookup

The archive carries a field that reads every published entry. There is no search server behind it: the matching happens in the browser, and the words typed are never sent anywhere. The engine (opens in a new tab) is Rust compiled to WebAssembly, and the same module writes the files at build time and reads them back in the tab, so one tokenizer decides what a word is on both sides.

A build writes three files, and only one of them is fetched whole. It holds the dictionary, the entry rows and the offsets, and it starts downloading when the field takes focus rather than when the first key lands, so the wait sits in front of the first keystroke instead of after it. The other two are read a range at a time: the posting lists of the words actually typed, and one window of reading text for each result printed. Growing a word one letter at a time reads nothing it already holds.

• Every word has to land. The last one is still being typed, so it reaches every ending of itself; the rest are taken as written.
• Order comes from BM25, with a bonus for words that stand next to each other in the entry. A line break parts two words the reader never saw together, so a phrase is not found across the seam between a caption and the paragraph beside it.
• A word the ledger does not hold is answered by the nearest one it does, counting two characters the wrong way round as a single slip. Both words are named under the count, because a name typed deliberately should not be quietly overruled.

Two failures are worth stating. Without JavaScript the field is hidden and the archive list stands on its own. And the addresses of all three files carry a digest of the index, so a page left open across a deploy finds its next range missing and says the ledger has been reprinted, rather than quoting text whose offsets belong to another build.