Problem
TextRenderer.LayoutParagraphWithWidth (both Skia and ImageSharp) measures text at the word level: each run's text is split via SplitIntoWords into individual word-or-whitespace fragments, and each is measured independently with font.MeasureText(word). Fragment widths are summed to get the line width, and fragments are rendered sequentially at currentX += fragment.Width.
This loses cross-word shaping — kerning pairs and bearing adjustments that a shaper would apply to "abc def" as a single run are gone once it becomes ["abc", " ", "def"]. Word measurement also diverges further when the source document stores adjacent characters as separate runs (e.g. cover-letters/02's contact line has 9 runs: "+91 915 5894669", " ", "|", " ", "www.interestingsite.com", " ", "|", " ", "manasi@example.com").
Symptom
Visible extra whitespace before specific characters — most obvious around | separators. Measured pixel width of the contact-info line in cover-letters/02:
- Expected: 616px total, ~14px gap before each
|
- Skia render: 633px total, 27px gap before the first
|, 18px before the second
The drift is non-uniform: it accumulates at fragment boundaries where shaping would normally tighten positions.
Proper fix
Switch the layout pipeline to shape entire lines:
- Skia: use
SKShaper (via SkiaSharp.HarfBuzz) to shape the full line's glyphs and pull per-character advances. Keep fragment identity (for wrapping, styling, underline etc.) but source their widths from the shaped glyph advances instead of font.MeasureText on isolated words.
- ImageSharp:
TextMeasurer.MeasureAdvance already handles shaping when called on the full string. Pre-measure each run's full text in context, then split the measured result into word slices using glyph clusters/char indices.
Line wrapping needs to still operate on word boundaries, but each word's width comes from its position inside the shaped line, not an isolated measurement.
Why this is a standalone task
An earlier attempt to retrofit shaping via post-layout fragment merge caused 158 regressions out of 280 scenarios. Every existing results_{skia,imagesharp}#page_*.verified.png was generated against the old word-by-word measurement, so any measurement change — even one that's objectively closer to Word's output — shifts pixel positions enough to fail the ErrorMetric.Absolute threshold on scenarios that were previously passing.
That means this fix must be committed together with fresh baselines for all affected scenarios.
Estimate
- Skia implementation (swap to
SKShaper, rewire LayoutParagraphWithWidth / LayoutParagraph / LayoutParagraphForMeasurement, update RenderFragment to draw via shaped glyphs or sub-strings): ~1 day
- ImageSharp implementation (mirror of above using
TextMeasurer): ~0.5 day
- Justified-text gap handling (
extraSpacePerGap in RenderParagraph) needs re-derivation from shaped advances: ~0.5 day
- Table cell layout (
RenderParagraphInBounds uses the same path): verify/fix: ~0.5 day
- Re-baselining ~280 scenarios: force-accept, visually spot-check the categories (cover-letters, resumes, business-plans, etc), compare error metrics against expected to make sure the new baselines are not worse overall: ~1 day
- Regression chase for anything that breaks in non-obvious ways (RTL, bullets, soft hyphens, inline images mixed with text): ~0.5–1 day
Total: ~4 days of focused work, landed as a single commit to keep the baseline refresh atomic.
Affected files (primary)
src/Morph.Skia/Rendering/TextRenderer.cs — LayoutParagraphWithWidth, LayoutParagraph, LayoutParagraphForMeasurement, RenderParagraph, RenderFragment, SplitIntoWordssrc/Morph.ImageSharp/Rendering/TextRenderer.cs — same set of methodssrc/Morph.Skia/Rendering/RenderContext.cs — font creation (may need shaper handle)src/Morph.ImageSharp/Rendering/RenderContext.cs — MeasureText helpersrc/Tests/Inputs/**/results_*.verified.png — baselines to refresh
Problem
TextRenderer.LayoutParagraphWithWidth(both Skia and ImageSharp) measures text at the word level: each run's text is split viaSplitIntoWordsinto individual word-or-whitespace fragments, and each is measured independently withfont.MeasureText(word). Fragment widths are summed to get the line width, and fragments are rendered sequentially atcurrentX += fragment.Width.This loses cross-word shaping — kerning pairs and bearing adjustments that a shaper would apply to
"abc def"as a single run are gone once it becomes["abc", " ", "def"]. Word measurement also diverges further when the source document stores adjacent characters as separate runs (e.g. cover-letters/02's contact line has 9 runs:"+91 915 5894669"," ","|"," ","www.interestingsite.com"," ","|"," ","manasi@example.com").Symptom
Visible extra whitespace before specific characters — most obvious around
|separators. Measured pixel width of the contact-info line incover-letters/02:||, 18px before the secondThe drift is non-uniform: it accumulates at fragment boundaries where shaping would normally tighten positions.
Proper fix
Switch the layout pipeline to shape entire lines:
SKShaper(via SkiaSharp.HarfBuzz) to shape the full line's glyphs and pull per-character advances. Keep fragment identity (for wrapping, styling, underline etc.) but source their widths from the shaped glyph advances instead offont.MeasureTexton isolated words.TextMeasurer.MeasureAdvancealready handles shaping when called on the full string. Pre-measure each run's full text in context, then split the measured result into word slices using glyph clusters/char indices.Line wrapping needs to still operate on word boundaries, but each word's width comes from its position inside the shaped line, not an isolated measurement.
Why this is a standalone task
An earlier attempt to retrofit shaping via post-layout fragment merge caused 158 regressions out of 280 scenarios. Every existing
results_{skia,imagesharp}#page_*.verified.pngwas generated against the old word-by-word measurement, so any measurement change — even one that's objectively closer to Word's output — shifts pixel positions enough to fail theErrorMetric.Absolutethreshold on scenarios that were previously passing.That means this fix must be committed together with fresh baselines for all affected scenarios.
Estimate
SKShaper, rewireLayoutParagraphWithWidth/LayoutParagraph/LayoutParagraphForMeasurement, updateRenderFragmentto draw via shaped glyphs or sub-strings): ~1 dayTextMeasurer): ~0.5 dayextraSpacePerGapinRenderParagraph) needs re-derivation from shaped advances: ~0.5 dayRenderParagraphInBoundsuses the same path): verify/fix: ~0.5 dayTotal: ~4 days of focused work, landed as a single commit to keep the baseline refresh atomic.
Affected files (primary)
src/Morph.Skia/Rendering/TextRenderer.cs—LayoutParagraphWithWidth,LayoutParagraph,LayoutParagraphForMeasurement,RenderParagraph,RenderFragment,SplitIntoWordssrc/Morph.ImageSharp/Rendering/TextRenderer.cs— same set of methodssrc/Morph.Skia/Rendering/RenderContext.cs— font creation (may need shaper handle)src/Morph.ImageSharp/Rendering/RenderContext.cs—MeasureTexthelpersrc/Tests/Inputs/**/results_*.verified.png— baselines to refresh