- Shipped
- August 26, 2026 at 4:22 AM UTC
- Author
- kamo
- Commit
- aa02328
Both faces are preloaded on every route, so the format is paid on every first visit. 134,132 B of WOFF became 112,252 B of WOFF2 — 21,880 B, a 16% cut. That is smaller than the ~30% a WOFF2 conversion is usually quoted at, and the reason is worth writing down so nobody re-does this expecting more: the 30% figure is against uncompressed TTF. WOFF is already zlib-compressed, so what WOFF2 adds here is brotli plus the glyf/loca transform, not compression where there was none. Converted by decoding each WOFF container back to bare sfnt and re-encoding. Verified lossless by round-tripping the result: all 17 tables present in both faces, and `fvar` (the variable axes — these are variable fonts carrying the whole 100-900 range), `name` and `STAT` byte-identical. The build emits both faces and `next/font/local` accepts them. The preload stays. Dropping it was suggested during the audit and is a real regression: Geist Mono would fall back to the generic monospace stack on first paint. The converter was installed outside the project (`npm --prefix /tmp`), so package.json and the lockfile are untouched — worth doing that way in a repo where several sessions are editing at once. WOFF2 is universal in every evergreen engine since 2016 and ~98.5% globally; this app runs Next 16 and React 19, which do not support anything older. The .woff files are removed rather than kept as a fallback, since a second `src` would have to be declared and would only ever serve a browser the framework already refuses. Verified: production build succeeds and emits both faces; tsc --noEmit clean; all 10 guards pass.