Font-family matching (match-font) + cmap parse-once cache

[ynniv]

Two additions from driving a web engine (weft) against scribe — a new consumer-facing API and a performance fix that turned out to be load-bearing for text-heavy pages.

1. match-font — CSS font-family matching (5edbe15)

Adds a match-font (family-list &key weight style) → font service so a consumer can resolve a CSS font-family stack + weight + style to a concrete face, with no NIL returns (graceful fallback chain ending at Liberation Sans Regular).

• Ships the metric-compatible Liberation face set (Sans/Serif/Mono × Regular/Bold/Italic/BoldItalic, + SIL OFL license) under fonts/.
• Case-insensitive family mapping (generics + Arial/Helvetica/Verdana/Times/Georgia/Courier/… → the right family), weight ≥ 600 → Bold, italic/oblique → Italic; lazy-open + cached.
• Exports font-ascent / font-descent / font-line-gap (hhea) so a consumer can compute line-height: normal from real metrics.

Verified against a browser: match-font('("Verdana" "sans-serif")) measures a fixed string at 420.7px, matching Chromium's Liberation-Sans substitution to the pixel.

2. Parse the cmap once, cache it on the font (b6f0012)

font-glyph-index rebuilt the entire cmap hash table on every codepoint lookup — reading the whole subtable (thousands of entries) to map a single character. On a consumer that measures every word repeatedly (a web engine's table auto-layout), this is quadratic and dominates wall time.

Fix: parse the best subtable once into a new lazy %cmap slot (mirroring the existing %cff/%fvar/%gdef caches); lookups are now a single gethash. Behavior-identical — same subtable, same mappings, glyph indices unchanged.

Impact measured downstream (weft rendering a 40 KB Hacker News page): 7.5 s → 0.18 s render, layout 6.9 s → 20 ms (345×), output byte-for-byte unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JA1saK7BkgavHNeurmK65k

[ynniv]

Merged via rebase (linear history preserved; two clean commits). Review notes for the record:

cmap cache — verified behavior-preserving: swept 34,816 codepoints (ASCII, BMP, and non-BMP incl. format-12 ranges) across LiberationSans/Serif/Mono + DejaVuSans, comparing the cached font-glyph-index against a fresh re-parse of the pre-cache logic on a separate font instance. 0 mismatches. Subtable-selection order is byte-identical to the old code; the empty-cmap path now caches an empty hash-table (parsed once, still returns 0) rather than re-parsing forever. Cache is properly memoized (font-cmap returns eq object across calls). Fonts are immutable after open-font, so per-font caching is sound.

match-font — never returns NIL across exact/unknown/empty-list/string-arg/generic/DejaVu inputs; distinct objects for regular vs bold; exports match-font/font-ascent/font-descent/font-line-gap all present; metrics plausible (upem 2048, ascent 1854, descent -434, lineGap 67). All 12 Liberation faces + OFL license committed (not gitignored) and load via open-font.

Suite — inspect/run-all.sh: 10/10 gates PASS (table x5, cmap x3, shape vs HarfBuzz x2); clean compile, no warnings. CI gates also green.

Non-blocking follow-up: match.lisp resolves DejaVu from inspect/corpus/ (test-only data) via *corpus-dir*. Not a blocker — match-font degrades to Liberation and never errors if the corpus is absent from a distribution — but shipping code depending on the test corpus is worth revisiting if DejaVu is meant to be a first-class shipped face.