paint: shape required-ligature ("rlig") text so icon web fonts paint

[tannevaled]

Summary

• go.dev/blog's own nav dropdown carets are Google's Material Icons ligature words (<i class="material-icons">arrow_drop_down</i>), substituted by the font's OpenType GSUB rlig (Required Ligatures) feature into one pictogram glyph. @font-face now loads that real font (css: named font families, and @font-face (1/3) #233/Fetch and register a document's own @font-face files (2/3) #234/@font-face reaches a font service at last (WOFF2) #236), but Measure/drawText only ever did a per-rune cmap+advance lookup — so the word's individual, real (non-.notdef) letter glyphs painted nothing at all, instead of the one substituted icon a real browser shows.
• Measure (paint/fonts.go) and drawText (paint/paint.go) now shape each run through Face.Shape(text, "rlig") and walk the resulting glyph indices (AdvanceIndexUnits/GlyphMaskIndex) instead of raw runes. rlig is the OpenType-spec-mandated always-on ligature feature (its own registry entry: "Control of this feature should not generally be exposed to the user"), unlike liga, which a real browser disables under non-zero letter-spacing — confirmed via MDN, and deliberately not modeled this round (see FIDELITY.md's "Known gaps").
• Verified none of the four bundled prose fonts (Inter, Lora, Go Mono, DejaVu Sans) declare any rlig lookups, so this is isolated to fonts that actually require it — zero risk to existing bundled-font text.
• Corrects FIDELITY.md's own stale "Known gaps" entry, which had said a fix would be "genuinely comparable in scope to Shadow DOM" based on a @font-face-fetching capability and a go-opentype/shape package that don't reflect the current code (@font-face shipped separately; the shaping API landed directly on opentype.Face, no separate package).

Test plan

• go build ./..., go vet ./..., go test ./... all green
• Coverage floors held (paint back to its 100.0% ratchet floor)
• New tests in paint/ligature_test.go, using the real Material Icons WOFF2 (Google, Apache License 2.0) as a fixture: Measure("arrow_drop_down") collapses to one glyph's advance; drawText paints visible ink for it; an uncovered rune still measures/paints nothing (the .notdef guard)
• Git-stash-confirmed: reverting paint/fonts.go/paint/paint.go makes the two new positive tests fail exactly as expected (unshaped 15-glyph width; no ink painted)
• Full 10-page bench comparison run: all pages within the same small run-to-run noise band already established for this live comparison; the fix's real effect (a few dozen pixels of one icon glyph) is too small a share of a full page to move aggregate SSIM outside that noise — confirmed instead by direct visual inspection of the isolated repro and a live re-render of go.dev/blog itself, both now showing a real triangle glyph where the nav previously rendered nothing

🤖 Generated with Claude Code