Make Line::items() glyph-run iteration linear and cluster-free

[nicoburns]

LLM Contributions: This PR was written with Devin, which also ran the benchmarks below.

Line::items() (the positioned GlyphRun iterator used by every renderer, e.g. Blitz/Vello) rebuilt run.visual_clusters().flat_map(|c| c.glyphs()).skip(self.glyph_start) on every next(). That is O(glyphs × glyph runs) per run, and since the parley_engine refactor each skipped element is a by-value Cluster (Run + Atom + Grapheme) plus a large composed glyph-iterator state, so this path went from ~200 ms to ~4 s of a Blitz layout of the Barack Obama Wikipedia page (see benchmark comment on DioxusLabs/blitz#832).

Two independent fixes, as separate commits so they can be measured separately:

1. Cursor (28ea7e7): GlyphRunIter keeps a peekable cluster cursor for the current run and resumes from it instead of restarting + skip. Removes the quadratic term.
2. Reference / slice (5a2c020): GlyphRunIter no longer materialises Clusters or calls Cluster::glyphs(). A small AtomCursor walks the run's ShapedSlice atoms (atoms_start()/atoms_end(), prev() for RTL), reads the style index from atom.characters()[0], and sums glyph counts/advances from shaped_cluster_glyphs() plus EffectiveSpacing::gaps(atom) — the same quantities Run::glyphs_in produces, without constructing the composed iterator. Glyph runs always end on atom boundaries (style is per character, Cluster::style_index already used the atom's first character), so resuming at atom granularity is exact.

Semantics (style_index, glyph_start, glyph_count, offset, advance, RTL order, letter/word spacing, justification, ligatures/glyph-less clusters) are unchanged; GlyphRun::glyphs() still uses glyphs_in(...).skip(glyph_start).take(glyph_count). The Clusters iterator is untouched for other consumers.

Also:

• parley_bench: new Styled Items benchmark that walks line.items() / positioned_glyphs() of a pre-built styled layout. None of the existing benchmarks executed GlyphRunIter::next, which is why this regression was invisible to Tango.
• parley/src/tests/test_glyph_runs.rs: checks items() against a naive split of visual_clusters().flat_map(glyphs) (style, glyph ids, advances, offsets) for latin/arabic/japanese/mixed text × spacing on/off × start/justify.

cargo fmt --all --check, cargo test --workspace, cargo clippy --workspace --all-targets -- -D warnings pass.

Benchmarks

Variants (all include the Styled Items bench commit f224cc7 so Tango can compare it):

	Revision	Description
base	f224cc7	main (83acf8f) + bench only
cursor	28ea7e7	fix 1 only
ref-only	scratch (not pushed): 5a2c020 with the cursor disabled — AtomCursor::new(run) per next() and skipping glyph_start glyphs atom by atom	fix 2 only
both	5a2c020	this PR

Tango, parley_bench Styled Items (<variant>/target/benchmarks/main compare <base>/target/benchmarks/main -t 1 -o; columns [base ... variant])

Benchmark	cursor	ref-only	both
arabic 20 chars	[12.1 us ... 6.2 us] -49%	[11.5 us ... 1.6 us] -86%	[11.4 us ... 886 ns] -92%
latin 20 chars	[11.6 us ... 5.9 us] -50%	[11.8 us ... 1.6 us] -86%	[11.6 us ... 713 ns] -94%
japanese 20 chars	[11.1 us ... 5.9 us] -47%	[11.3 us ... 1.6 us] -86%	[12.2 us ... 971 ns] -92%
arabic 1 paragraph	[67.5 us ... 32.8 us] -51%	[71.3 us ... 9.7 us] -86%	[66.8 us ... 5.0 us] -93%
latin 1 paragraph	[51.7 us ... 29.1 us] -44%	[56.8 us ... 7.7 us] -87%	[52.2 us ... 3.1 us] -94%
japanese 1 paragraph	[125.3 us ... 62.3 us] -50%	[117.5 us ... 16.4 us] -86%	[116.7 us ... 9.1 us] -92%
arabic 4 paragraph	[275.6 us ... 133.7 us] -52%	[268.4 us ... 37.0 us] -86%	[266.5 us ... 20.5 us] -92%
latin 4 paragraph	[207.2 us ... 116.3 us] -44%	[211.4 us ... 28.5 us] -87%	[213.5 us ... 13.0 us] -94%
japanese 4 paragraph	[161.5 us ... 80.5 us] -50%	[163.3 us ... 22.8 us] -86%	[171.1 us ... 13.4 us] -92%

All other benchmarks (Default Style, Styled, Word + Letter Spacing, Repeated Justification) don't call items(); they moved within ±3% in either direction across the three comparisons (noise/code layout).

Blitz, Barack Obama Wikipedia page, headless, 1280 px (same harness/inputs as the DioxusLabs/blitz#832 report)

Blitz main (1884860, with the API-compat edits needed for Parley main) built --release with [patch] pointing at each Parley worktree; 3 warm-ups + 20 iterations per run, 3 interleaved rounds (base cursor ref both ×3), medians of the per-round medians:

Variant	First parse+style+layout median / p90 (ms)	Relayout median / p90 (ms)	Peak RSS (MB)
base (Parley main)	1051 / 1109	929 / 986	186.8
cursor	473 / 510	352 / 371	186.7
ref-only	303 / 328	183 / 194	186.7
both (this PR)	263 / 282	151 / 172	186.8

Per-round relayout medians: base 929/926/946, cursor 352/348/353, ref-only 185/181/183, both 144/151/156. For reference, Blitz main with crates.io Parley 0.11.1 measured 147 ms relayout on the same machine, so this restores the pre-refactor cost of the renderer's item walk.

Interpretation: the slice walk alone buys more than the cursor alone (the constant per skipped element was ~18× the quadratic-term exposure at Wikipedia-paragraph glyph-run counts), but only both together make the iteration O(n) with cheap elements; the remaining line.items() cost in the Blitz profile is now negligible.

Changelog

Changed

• Line::items() no longer re-scans a run's glyphs for every glyph run it yields, and no longer builds Clusters to do so; iterating a line's positioned items is now linear in the number of glyphs.

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/3a0b8cc210c041d8955c841fb2a76a8e
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/3a0b8cc210c041d8955c841fb2a76a8e?variant=devin-insiders
Requested by: @nicoburns

[staging-devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring