This site against Stripe, Linear, Column and Increase

The homepage’s 50 scripts sit before </body> and carry no defer, deliberately. defer pushed them past First Contentful Paint, where Lighthouse starts counting blocking time. A/B on production: with defer, 300 ms and 83; without, 100 ms and 90. It cost 0.3 s of FCP. TBT weighs 30 in the score, FCP weighs 10.

A build script writes titles, descriptions, canonicals, Open Graph cards and JSON-LD from the filesystem, so they cannot drift from the pages they describe. I run six deterministic gates before every deploy, and a regression fails it.

These three choices are for people who use a screen reader, people who move through a page by keyboard, and people who tap with a thumb on a phone. Tab panels sit on <section> because ARIA in HTML gives <article> six allowed roles and tabpanel falls outside them. Touch targets reach 24 px with a pseudo-element. A transform would push the button past its box and overlap its neighbor. I mark inactive slides inert, so the whole block leaves the tab order. Under aria-hidden they stay focusable and unannounced. The count below comes from axe, an accessibility scanner.

The dot engine behind this site’s illustrations reads only light and dark, so every color channel in its source images was stored for nothing and the decoder threw it away. What it samples is Rec.601 luminance, the black-and-white brightness of a pixel. I re-encoded them to 8-bit greyscale, reproducing the engine’s alpha-threshold rule, and checked the luminance byte for byte. 111 files, 7.87 → 2.81 MB. Vendor logos draw at 16–20 px; one was a 204 KB multi-resolution .ico.

Fonts came from two external origins whose stylesheets pointed at a third, so first paint waited on a four-hop chain: 258 ms to the CSS, 813 ms to the last font. I host and subset all seven faces. One rule uses Newsreader at italic 500. It shipped as a variable font with two axes.

Six of the desktop run’s seven long tasks were the same file: the full-viewport WebGL shader behind the hero. A machine with no GPU has to draw that shader on its processor, which leaves little room for anything else, so those machines now get the static gradient instead. The page checks WEBGL_debug_renderer_info and makes the switch on SwiftShader, llvmpipe and Microsoft Basic Render. Rasterizing in software costs 65 ms a frame on the CPU, so a GPU-less machine got 15fps and a pegged core.