Fix glyph cache poisoning across renders; add CLAUDE.md and handoff.md

fontkit caches glyphs per font for the whole process and pdfkit embeds composite
components with no code points, so setting ':' before '.' in Poppins corrupted every
later render in the same process (text layer showed '1,200;00', line breaks shifted).
primeFontGlyphs() now creates the component glyphs with their cmap code point at font
registration, awaited by renderCore and browser init. Regression test and Quirk 13 added;
the warm-up workarounds in the template tests are removed.
This commit is contained in:
2026-10-04 13:11:40 +05:30
parent 76533ad2c7
commit 63854dd960
9 changed files with 308 additions and 25 deletions
+72
View File
@@ -31,6 +31,78 @@ export function registerBundledFonts(base: string): void {
Font.registerHyphenationCallback((word) => [word]);
}
/** The slice of a fontkit TTF font the priming touches (fontkit's types do not declare it). */
interface FkFont {
numGlyphs: number;
characterSet: number[];
glyphForCodePoint(cp: number): unknown;
_getBaseGlyph(id: number): { _decode?(): { components?: Array<{ glyphID: number }> } | null };
_glyphs: Record<number, unknown>;
_cmapProcessor: { lookup(cp: number): number };
}
const isPrivateUse = (cp: number) => cp >= 0xe000 && cp <= 0xf8ff;
/**
* Creates, through the cmap and so with their code point, every glyph that a composite glyph uses as a
* component. Not the whole cmap: a glyph that layout substitutes in (Inter's contextual hyphen between digits)
* is mapped from a private-use code point, and creating it from there would give it a code point the text
* never had. Components that only a private-use code point reaches are left alone.
*/
function primeComponents(font: FkFont): void {
const components = new Set<number>();
// _getBaseGlyph caches what it creates, with no code points; scan into a scratch cache so that does not happen.
const cache = font._glyphs;
font._glyphs = {};
try {
for (let id = 0; id < font.numGlyphs; id++) {
const decoded = font._getBaseGlyph(id)._decode?.();
for (const c of decoded?.components ?? []) components.add(c.glyphID);
}
} finally {
font._glyphs = cache;
}
const lowest = new Map<number, number>();
for (const cp of [...font.characterSet].sort((a, b) => a - b)) {
if (isPrivateUse(cp)) continue;
const id = font._cmapProcessor.lookup(cp);
if (components.has(id) && !lowest.has(id)) lowest.set(id, cp);
}
for (const cp of lowest.values()) font.glyphForCodePoint(cp);
}
let primed: Promise<void> | null = null;
/**
* Loads every registered face and gives the component glyphs of its composites their code points. fontkit
* caches glyph objects per font for the whole process, and pdfkit creates a composite's components at embed
* time with no code points: Poppins' colon contains the period, so a document that sets ':' before any '.'
* would leave the period glyph with codePoints [] for every later document (amounts extracted as "1,200;00",
* textkit's line breaks slipping). Creating those glyphs here first makes a render independent of the renders
* before it. Memoised; await it before the first render.
*/
export function primeFontGlyphs(): Promise<void> {
if (!registered) return Promise.resolve();
if (!primed) {
primed = run().catch((err) => {
primed = null;
throw err;
});
}
return primed;
async function run(): Promise<void> {
for (const fam of FONT_MANIFEST.families) {
for (const s of familySources(fam)) {
const descriptor = { fontFamily: fam.family, fontWeight: s.fontWeight, fontStyle: s.fontStyle };
await Font.load(descriptor);
const data = (Font as unknown as { getFont(d: object): { data: FkFont | null } }).getFont(descriptor).data;
if (data) primeComponents(data);
}
}
}
}
const MIN_FONT_BYTES = 10 * 1024;
function hasTrueTypeMagic(b: Uint8Array): boolean {