nx_ttf_fontlib.nx

nx_ttf_fontlib.nx -- the sovereign TrueType FONT ENGINE as a LIBRARY (the browser's text renderer). Canonical home of the TTF read path first proven in the nx_ttf_raster demo (which has a main() and so cannot be imported): sfnt dir -> head/cmap/loca/glyf/hhea/hmtx(+kern), cmap format-4 (BOTH the delta path and the idRangeOffset/glyphIdArray path), glyf outline decode (flags/short/same/repeat), COMPOSITE glyphs (component offsets + F2Dot14 scale, bounded recursion), quadratic Bezier flattening, NONZERO- WINDING scan conversion with SSxSS supersampled ANTI-ALIASED coverage, a per-(char,size) COVERAGE CACHE (each glyph rasterizes once per size), REAL per-glyph advances from hmtx, pair KERNING from the 'kern' table (format 0, binary search), a mild contrast curve (stem darkening at small sizes), and an RGBA-framebuffer ALPHA-BLEND draw (nx_framebuffer px = R,G,B,A). 2026-09-02 (BR4 first half, /compare/browser rung "Real font files + shaping"): EVERY METRIC IS NOW READ FROM THE FONT, NEVER ASSUMED. The first cut hard-coded unitsPerEm=1000 and ascender=760 -- true of our own emitted nishi_sans.ttf and FALSE of every professional font (Liberation Sans, Arial, Noto: 2048), so a real face would have rasterized at HALF size with its advances halved, and nothing would have failed loudly. unitsPerEm comes from head, ascender/descender/lineGap from hhea, the cell height is DERIVED from (ascender - descender), and the browser's 1000-em advance unit is a conversion at the API edge (tf_adv_em1000), not an assumption inside the rasterizer. A negative left side bearing ('j', '_', 'w' in Liberation) is carried as the record's left margin so no ink is clipped, and ink wider than the advance ('f', 'k') widens the cell instead of being cut at the advance. Fonts proven: our own emitted web_assets/nishi_sans.ttf (96 glyphs, upem 1000) and Liberation Sans (SIL-OFL, upem 2048, 908 kern pairs, 1,076 composite glyphs). No FreeType, no GDI, no DirectWrite. license_tier: ORIGINAL