Fix text kerning, HiDPI glyph padding, and invisible-glyph rasterization

[pusewicz]

Summary

A review of src/cute_draw.cpp's text rendering path turned up three independent bugs. All are fixed here.

1. Kerning was not applied to the text layout cursor

advance_to_next_glyph in s_draw_text only advanced x by glyph->xadvance. The only place kerning was used was a local offset added to a single glyph's quad in the render block, which:

• Didn't accumulate — every glyph past the first carried none of the preceding kern adjustments.
• Was gated behind if (render || markups), so the measurement pass (cf_text_size, word-wrap) never saw it, making returned sizes and wrap points disagree with what was actually drawn.

cf_font_get_kern also looked up the wrong key, so it usually returned 0 outright: the kerning table is built from stbtt_GetKerningTable, keyed by glyph index, but the lookup queried it with raw codepoints.

Fixing the lookup with stb's stbtt_GetCodepointKernAdvance (which converts codepoint -> glyph index and also reads GPOS) still returned 0 for classic pairs like "AV"/"To" on this project's bundled Calibri. A standalone probe against the embedded font data showed why: Calibri has a populated legacy kern table (26,706 entries, real values e.g. "To" = -182 units) and a GPOS table, but stb's minimal GPOS parser doesn't support this font's GPOS lookup layout, and stb only consults the legacy kern table when there's no GPOS table at all. The fix now tries GPOS first and falls back to the (correctly glyph-index-keyed) legacy table when GPOS reports nothing.

s_find_end_of_line (word-wrap) also now folds kerning into its width accounting, and the cursor correctly skips kerning for the first glyph of the string (cp_prev starts at 0 -- thanks for catching that in the follow-up commit).

2. HiDPI glyph padding was off by pixel_scale

The glyph quad's border pad was a flat V2(1,1) logical unit, meant to cover spritebatch's 1-pixel atlas border. That border is 1 device pixel, while glyph bitmaps rasterize at pixel_scale resolution and q0/q1 are logical units -- so on Retina displays the pad overshot the border, stretching every glyph quad slightly past its actual coverage (worse for smaller glyphs). Fixed by dividing the pad by app->pixel_scale, matching how q0/q1/xadvance are already scaled elsewhere in the same function.

3. Every space character was rasterized, given an atlas slot, and drawn as a blank quad

glyph->visible |= w > 0 && h > 0 always evaluated true (w/h include border padding, which is never zero), silently overriding stbtt_IsGlyphEmpty's correct verdict. Every invisible glyph (spaces, etc.) ended up rasterized, cached into the atlas, and pushed as a blank draw quad on every occurrence. Beyond fixing the flag, invisible glyphs now skip s_render's rasterization and atlas registration entirely -- layout metrics (xadvance, q0/q1) are still computed since those are needed regardless of visibility. A new rendered flag on CF_Glyph replaces the old image_id-based cache-hit check in cf_font_get_glyph, since invisible glyphs never get an image_id.

Changes (src/cute_draw.cpp, src/internal/cute_font_internal.h)

• cf_font_get_kern: try stbtt_GetCodepointKernAdvance first, fall back to a glyph-index lookup into the legacy kern table.
• s_draw_text: accumulate kern into the x cursor before positioning each glyph (horizontal layout only, skipped right after a line break or for the first glyph), unconditionally so measurement stays consistent with rendering. Removed the now-redundant per-glyph kern offset on the quad.
• s_find_end_of_line: word-wrap width accounting now includes kerning too.
• Glyph quad padding now divides by app->pixel_scale to match the device-pixel atlas border.
• s_render/cf_font_get_glyph: invisible glyphs skip rasterization and atlas registration; CF_Glyph::rendered tracks "already processed" instead of image_id.

Test plan

• cmake --build build -- clean, no new warnings
• ./build/tests -- 173/173 passing
• Visual check via samples/font_debug.cpp against the bundled Calibri: kerning readout for "To" went from 0.00 to -2.31 at font size 26; before/after screenshot crop confirms the o glyph is now pulled tight under the T
• Debug print confirmed space-glyph behavior directly: visible=0 rendered=1 image_id=0 xadvance=5.88 -- no rasterization or atlas slot allocated, but layout spacing is unaffected

https://claude.ai/code/session_011V2maY4DG1DHscsMEHy6Bf

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[Copilot]

Pull request overview

This PR fixes text layout kerning so kerning affects the layout cursor (measurement and wrapping) rather than a per-glyph render-only offset, and corrects kerning data lookup by using stb’s codepoint kerning API with a fallback to the legacy kern table keyed by glyph index.

Changes:

• Update cf_font_get_kern to prefer stbtt_GetCodepointKernAdvance (GPOS) and fall back to the legacy kerning map using glyph indices.
• Update s_draw_text to accumulate kerning into the horizontal cursor so layout/measurement matches rendering.
• Update s_find_end_of_line to include kerning in wrap width accounting.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.