Release 0.9.0

CHANGELOG.md

@@ -1,3 +1,32 @@
+# 0.9.0
+
+Released on Friday, October 30 2026.
+
+- Fixed `-webkit-line-clamp` counting lines per inline run instead of across the nested blocks of the clamped box (e.g. paragraphs in a clamped card), which now ends after its last line
+- Fixed gaps with font- or viewport-relative units (e.g. Tailwind's `gap-4`, i.e. `gap: 1rem`) being ignored in grid and flex containers
+- Fixed images (and other replaced elements) with `border-radius` not being clipped to its curve (e.g. a `rounded-full` avatar painted square)
+- Fixed `line-height` lengths (e.g. `16px`, or `1rem` as in Tailwind's `text-xs`) being ignored in favor of the inherited (or default) line height
+- Fixed inline-level flex items and non-stretched grid items (e.g. a `<span>`) being as tall as their inline box instead of their line boxes, which let text overflow them at the bottom (e.g. tag pills with padding)
+- Fixed words breaking after a hyphen or slash in a box sized to fit them (e.g. a flex item with the max-content width of `c-sharp`), as the width came out of the box arithmetic a fraction of a pixel short
+- Added grid track alignment: `justify-content` and `align-content` (`start`, `end`, `center`, `space-between`, `space-around`, `space-evenly`, with `safe`) position the tracks of a grid whose tracks do not fill it
+- Added lines moving down past floats: a line whose first word does not fit beside the floats moves down to where one of them ends, until it fits or no float is beside it
+- Added `writing-mode` on flex, grid and table containers: a container whose writing mode differs from its parent's lays out its items along its own axes (e.g. a `vertical-rl` flex container runs its main axis downwards)
+- Added nested writing modes in the same orientation: a `vertical-lr` element in `vertical-rl` content (or the reverse) stacks its blocks in its own direction
+- Added `position: sticky` inside orthogonal flows (e.g. in a `vertical-rl` box), which used to stay static
+- Added anonymous table objects: stray `display: table-cell` boxes (e.g. `<div style="display:table-cell">` used for layout) form a row of an anonymous table, rows outside a table get one, and text or other content directly inside a table or row is wrapped in an anonymous row or cell instead of being dropped
+- Added percentage table heights (e.g. `table { height: 50% }`), resolved against a definite containing block height
+- Added `column-span: all` in multi-column containers: the spanner is laid out across all columns, and the content on each side of it is balanced separately
+- Added text shaping with HarfBuzz (via `SkiaSharp.HarfBuzz`): kerning and ligatures (e.g. `fi`, `fl`), Arabic joining and complex scripts such as Devanagari, measured and painted from the same shaped glyphs, so line breaking, alignment and table widths use the shaped advances
+- Added `font-feature-settings`, `font-kerning`, `font-variant-ligatures` and `font-variant-numeric` (e.g. `font-feature-settings: "liga" 0` turns ligatures off, `tabular-nums` selects tabular figures); a non-zero `letter-spacing` turns optional ligatures off
+- Added Noto Sans Devanagari as a bundled fallback font (its OFL license is included in the package), so Devanagari renders without an installed font
+- Changed right-to-left text in the display list: `DrawTextCommand.Text` is now in logical order with `IsRightToLeft` set, and shaping reverses and mirrors it (it used to be reversed and mirrored by layout)
+- Changed `text-overflow: ellipsis` to cut at glyph clusters (`ITextMeasurer.GetClusterBoundaries`, which defaults to grapheme clusters for custom measurers), so it never splits a ligature or conjunct
+- Improved SVG `text`/`tspan`, which is shaped the same way as HTML text, reordered by the bidi algorithm with `direction` as its paragraph direction (which also decides where `text-anchor: start` and `end` are), and in its `lang`
+- Added `lang` to text shaping, so fonts with language-specific forms use them (e.g. the Marathi forms of Noto Sans Devanagari for `lang="mr"`)
+- Added joining across element boundaries: an Arabic word split by an inline element (e.g. one letter in `<b>`) keeps its joined forms
+- Changed Canvas 2D `fillText` to use the context's `font` (family, size, weight, style) with the same fonts, shaping and bidi reordering as HTML text instead of the host's default typeface
+- Added fragmentation in multi-column containers: `break-before`/`break-after` (`column` forces a new column, `avoid`), `break-inside: avoid`, `orphans`/`widows` (lines of a paragraph are no longer left alone at the top or bottom of a column), `column-fill: auto|balance` with a definite height (balanced by default), and blocks without lines splitting across columns
+
 # 0.8.2
 
 Released on Sunday, October 25 2026.

docs/general/01-Basics.md

@@ -61,6 +61,10 @@ Inline content is laid out in line boxes, so nested inline elements keep their o
 Generic font families like `serif`, `sans-serif`, and `monospace` always map to fonts bundled with the renderer, so they render the same on every machine.
 Named families resolve only when installed on the host (or declared via `@font-face`), so prefer generic families when you need predictable output in visual tests.
 
+Text is shaped with HarfBuzz, so kerning and ligatures (`fi`, `fl`) apply, Arabic letters join, and complex scripts such as Devanagari form their conjuncts.
+`font-feature-settings`, `font-kerning`, `font-variant-ligatures` and `font-variant-numeric` select the OpenType features of the font; a non-zero `letter-spacing` turns optional ligatures off.
+The bundled fonts cover Latin, Greek, Cyrillic, Hebrew and Arabic (DejaVu) and Devanagari (Noto Sans Devanagari, used only as a fallback); other scripts need an installed font or an `@font-face` that covers them.
+
 ## Visual Testing
 
 The test suite uses PNG snapshots under `verification-assets/`. Missing baselines are created automatically unless strict mode is enabled with `ANGLESHARP_SNAPSHOT_STRICT=1`.

docs/tutorials/03-Questions.md

@@ -6,9 +6,13 @@ section: "AngleSharp.Renderer"
 
 ## Why do web-safe fonts look the same on my machine?
 
-The renderer maps generic families such as `serif`, `sans-serif`, and `monospace` to installed local fonts. If your system does not have distinct serif/sans/monospace faces available, Skia may fall back to similar-looking defaults.
+The renderer maps generic families such as `serif`, `sans-serif`, and `monospace` to fonts bundled with it (DejaVu), not to installed fonts, so they look the same on every machine.
 
-For stable results, install a font set with visibly different faces or use explicit family names that are present on the target machine.
+Named families (e.g. `Arial`) resolve only when installed on the host or declared via `@font-face`; otherwise the next entry of the `font-family` list is used.
+
+## Why does my text in another script show boxes?
+
+Characters the resolved font has no glyph for fall back to the next `font-family` entry, then the bundled fonts (Latin, Greek, Cyrillic, Hebrew, Arabic, Devanagari), then whatever the host has installed. For other scripts (e.g. Thai or CJK) declare an `@font-face` or install a font that covers them.
 
 ## Why does a snapshot fail when I add a new baseline?
 

src/AngleSharp.Renderer.Benchmarks/BASELINE.md

@@ -35,6 +35,10 @@ Against `AngleSharp` 1.8.1, `AngleSharp.Css` 1.1.2 (both via `PackageReference`)
 | **Layout only: `BuildDisplayList`** | **364 ms** | **194.14 MB** |
 | Full pipeline: layout + Skia rasterization + PNG encoding (`RenderToPng`) | 527 ms | 199.41 MB |
 
+> **Note (2026-10-28):** the tables above predate HarfBuzz text shaping and were measured on the
+> Apple M5. The shaping change was measured before/after on a different machine (see the last
+> History entry); re-run these tables on the M5 to refresh them.
+
 **Cumulative improvement over the original, pre-optimization baseline: ~9.5x faster, ~7.5x less
 allocation** on `BuildDisplayList` (net10.0: 2,415 ms / 1,456 MB → 255 ms / 194 MB). The
 net8.0 mean above carries a wide error margin in this particular short-job run (~285ms on a 364ms
@@ -206,3 +210,37 @@ BenchmarkDotNet refuses to run outside `Release` and warns loudly if it detects
 unoptimized build - always use `-c Release`. Update the tables above (and the date) after a
 meaningful optimization lands, so this file keeps tracking the current baseline rather than one
 frozen the day it was first written.
+
+### 2026-10-28, HarfBuzz text shaping (cost, plus a width cache in `SkiaTextMeasurer`)
+
+Measured on a different machine than the tables above, so only the before/after pair is
+comparable: AMD Ryzen 7 3700X (16 threads), Linux (WSL2), .NET SDK 10.0.100, `short` job.
+"Before" is commit `59b892a` (unshaped `canvas.DrawText`), "shaped" is HarfBuzz shaping in
+`SkiaTextShaping` without any caching, "shaped + cache" adds the per-thread width cache that
+`SkiaTextMeasurer.MeasureWidth` now keeps (keyed by text and `RenderFont`, cleared at 16,384
+entries).
+
+| `BuildDisplayList` | net10.0 Mean | net8.0 Mean | Allocated |
+|---|---:|---:|---:|
+| before | 423 ms | 401 ms | 119.9 MB |
+| shaped | 443 ms | 509 ms | 142.4 MB |
+| shaped + cache | 409 ms | 355 ms | 121.7 MB |
+
+| `RenderToPng` | net10.0 Mean | net8.0 Mean | Allocated |
+|---|---:|---:|---:|
+| before | 2,630 ms | 2,586 ms | 124.7 MB |
+| shaped | 2,923 ms | 2,988 ms | 148.1 MB |
+| shaped + cache | 2,867 ms | 2,963 ms | 127.3 MB |
+
+The short job's error margins are wide on this machine (up to +/-30 ms StdDev for layout), so the
+layout means are best read as "back to parity" with the cache; allocation is deterministic and
+ends up +1.8 MB (+1.5%). Painting keeps a real cost of roughly +9-15% on the full pipeline: every
+painted run is shaped once more when it is drawn (lines are mostly unique, so a paint-side cache
+would rarely hit).
+
+Nested trial layouts: a probe laying out 20 copies of inline-blocks nested 1-5 deep (each with a
+12-word sentence) in a 300px box. The number of `MeasureWidth` calls is identical before and after
+(trial layouts multiply it by ~2-2.5x per level: 1,280 / 4,020 / 9,760 / 21,500 / 41,946), and
+with the width cache the repeated calls cost a dictionary lookup instead of a shaping pass, so the
+time per layout did not grow (before: 17 / 36 / 89 / 160 / 154 ms, after: 19 / 45 / 76 / 133 /
+123 ms for depths 1-5; single runs of 5 layouts, so +/-20%).

src/AngleSharp.Renderer.Docs/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anglesharp/renderer",
-  "version": "0.8.2",
+  "version": "0.9.0",
   "preview": true,
   "description": "The doclet for the AngleSharp.Renderer documentation.",
   "keywords": [