[Text] Snap a right-to-left buffer split to the cluster boundary

[Gillibald]

What does the pull request do?

Fixes ShapedBuffer.Split producing two halves whose character and glyph coverage disagree when the split point falls inside a cluster of a right-to-left buffer. Downstream, this crashes the bidi reorderer with a NullReferenceException.

Second layer of a stack; based on #22065.

What is the current behavior?

A cluster straddling the split point keeps all of its glyphs in the leading half. This prevents breaking a cluster, which is correct; but the text is still cut at the requested offset. The leading half then claims fewer characters than its glyphs cover, and the trailing half claims characters whose glyphs it does not have. The left-to-right path (SplitAscending) already snaps the text boundary to the cluster; the right-to-left path did not.

// "أَبْجَدِيَّة": each Arabic letter carries its harakat in the same cluster,
// so the first cluster spans two characters.
var split = buffer.Split(1);
// split.First.Text.Length == 1, but split.First holds 2 glyphs

Both halves share the parent's cluster cache, which now describes neither of them, so everything built on it misleads in turn: FindTrailingCharCountWithinWidth returns a count that cuts a cluster, and splitting the half at that count returns a leading half holding every glyph and no trailing half at all. That null run reaches BidiReorderer.BidiReorder and throws.

What is the updated/expected behavior with this PR?

The text boundary follows the glyphs, so the halves stay consistent: no character and no glyph is lost or duplicated, and every glyph of the trailing half belongs to a character the trailing half owns. Should_Not_Split_RightToLeft_Cluster asserts exactly that. The Collapse crash disappears.

As with the left-to-right path, Split(n) may return a leading half slightly longer than n when n lands inside a cluster - that is the only way to keep a cluster whole.

How was the solution implemented (if it's not obvious)?

Checklist

• Added unit tests (if possible)?
• Added XML documentation to any related classes? (no new API; the existing SplitDescending remarks describe the boundary and were updated in place)
• Consider submitting a PR to https://github.com/AvaloniaUI/avalonia-docs with user documentation

Breaking changes

ShapedBuffer.Split(n) on a right-to-left buffer can now return a leading half longer than n when n falls inside a cluster, where it previously returned exactly n characters with a glyph set that did not match them. This mirrors the existing left-to-right behaviour. No other public API changes.

Obsoletions / Deprecations

None.

Fixed issues

No tracked issues; found while fixing #14011.

[MrJul]

LGTM!

[MrJul]

Well, it seems like the merge queue isn't working correctly with stacked PRs, even though there's a nice UI showing that it depends on the previous one in the queue. Let's wait for #22065 to get through.

[MrJul]

Note: the force-push above happened automatically with the stacked PRs. It also automatically dismissed my review.

[avaloniaui-bot]

You can test this PR using the following package version. 12.2.999-cibuild0068868-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[avaloniaui-bot]

You can test this PR using the following package version. 12.2.999-cibuild0068879-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]