[matmen]

This is step 2 of color emoji: GSUB shaping. Emoji sequences now render as a single glyph, and text fonts get their default ligatures.

Shaping (src/render/gsub.ts)

• Features: ccmp, rlig, liga, clig and calt, the ones browsers enable by default.
  • Records from all scripts and language systems apply, the same way the GPOS kerning already works.
• Lookup types:
  • 1 single, 2 multiple (an empty sequence deletes the glyph), 4 ligature
  • 5 contextual and 6 chained contextual, each in formats 1–3, including nested lookups
  • 7 extension
• Lookup flags: "ignore bases, ligatures or marks" and the mark attachment class, using the glyph classes from GDEF.
• Speed: each lookup remembers, per glyph, whether any of its subtables can start there. This keeps text with many lookups (Source Sans) as fast as before.

Font choice per grapheme (src/render/text.ts)

• Graphemes: a grapheme is a base character plus what extends it: combining marks, variation selectors, skin tones, the keycap, tags, ZWJ-joined characters, and pairs of regional indicators.
• Choosing the font: each grapheme comes from the first font that has all of its visible characters. If none has all of them, the font is chosen per character, as in renderText: fallback fonts and color emoji (COLR v0, CBDT) #75.
  • With FE0F, a font with a color glyph wins, so ❤️ is the red emoji heart and not Source Sans' black ❤.
  • With FE0E, a font without a color glyph wins.
• Runs: consecutive graphemes with the same font form a run, and the run goes through that font's GSUB. Invisible characters that no substitution consumed still take no room.

Loading and caching

• Fallback fonts are only loaded once a character needs them.
• Glyphs are kept per font, size and hinting across calls, with up to 16 sizes per font. This includes decoded and scaled emoji PNGs. A repeated call with Noto emoji took 6 instead of 49 ms; a line of Open Sans takes 0.3 instead of 1.1 ms.

Visible changes

• Emoji sequences are now single glyphs: 👍🏽, 🇩🇪, 1️⃣, #️⃣, 👨‍👩‍👧, 👨‍🦰, 🏳️‍🌈 and 🏴󠁧󠁢󠁳󠁣󠁴󠁿.
• Text fonts use ligatures: e.g. "office" now uses the ffi ligature, as in browsers. None of the existing targets contain ligatures, so they all still pass.
• Combining accents are still not positioned (that needs GPOS mark positioning). This was already the case before.

Tests

• These sequences must render as single glyphs: 👍🏽, 🇩🇪, 1️⃣, 👨‍👩‍👧.
• ❤️ with [sans, emoji] must come out as the red emoji.
• The emoji-cbdt and emoji-colr targets now include sequences.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJgRK5qpwJEbHPuFv6kgAQ