[web_paragraph] Avoid browser text metrics queries on re-layout and speed up first layout - #195122

…r rects

Each layout(width) also asked the span metrics for the selection of every block's trailing
whitespace. Those clusters are cached like the rest of the span, so the width now comes from
their rects as well; paragraph_relayout_test checks block and line widths against the span
metrics.

Perf: relayout_latin_long 880 -> 780 us (-11.4%, 95% CI [-13.4, -9.8], n=90/90) WIN
Perf: relayout_paint_latin_long 4773 -> 4718 us (-1.2%, 95% CI [-2.6, -0.0], n=90/90) NOISE
Perf: relayout_multistyle 705 -> 635 us (-9.9%, 95% CI [-10.6, -7.3], n=90/90) WIN
Perf: relayout_paint_multistyle 3468 -> 3473 us (+0.1%, 95% CI [-1.4, +1.0], n=90/90) NOISE
Perf: relayout_rtl_mixed 390 -> 360 us (-7.7%, 95% CI [-8.8, -6.3], n=90/90) WIN
Perf: relayout_paint_rtl_mixed 2645 -> 2645 us (+0.0%, 95% CI [-0.9, +0.7], n=90/90) NOISE
Perf: relayout_many_lines 1198 -> 1185 us (-1.0%, 95% CI [-1.9, +0.0], n=90/90) NOISE
Perf: relayout_paint_many_lines 4958 -> 4932 us (-0.5%, 95% CI [-2.3, +1.3], n=90/90) NOISE

Assisted-by: Claude Opus 5.5 (Medium)

…layout

Every layout(width) queried getActualBoundingBox for each new text block to fold its ink bounds
into the line and paragraph paint bounds, which only the painter reads. Line ink bounds are now
measured on first use and the paragraph paint bounds are computed on first read after a layout,
so a re-layout makes no ink queries and a paint pays for them once.

Perf: relayout_latin_long 797.5 -> 555 us (-30.4%, 95% CI [-31.9, -29.3], n=90/90) WIN
Perf: relayout_paint_latin_long 5098 -> 4748 us (-6.9%, 95% CI [-8.9, -3.7], n=90/90) WIN
Perf: relayout_multistyle 640 -> 440 us (-31.2%, 95% CI [-32.0, -30.6], n=90/90) WIN
Perf: relayout_paint_multistyle 3565 -> 3438 us (-3.6%, 95% CI [-4.7, -2.0], n=90/90) WIN
Perf: relayout_rtl_mixed 360 -> 255.1 us (-29.2%, 95% CI [-30.1, -27.8], n=90/90) WIN
Perf: relayout_paint_rtl_mixed 2627 -> 2590 us (-1.4%, 95% CI [-1.8, -0.4], n=90/90) NOISE
Perf: relayout_many_lines 1185 -> 877.5 us (-26.0%, 95% CI [-26.5, -25.4], n=90/90) WIN
Perf: relayout_paint_many_lines 4972 -> 5032 us (+1.2%, 95% CI [-0.9, +2.9], n=90/90) NOISE

Assisted-by: Claude Opus 5.5 (Medium)

TextWrapper.ellipsize stored the shaped ellipsis in TextLayout.ellipsisClusters, which nothing
cleared, so after a layout that truncated the text every later layout attached an ellipsis to
its lines, even at a width where the whole text fits, and re-measured it on every line.
wrapText now starts each layout without an ellipsis.

Assisted-by: Claude Opus 5.5 (Medium)

debugTextMetricsQueryCount counts measureText, getSelectionRects and getActualBoundingBox calls,
and paragraph_relayout_test asserts that re-layout at several widths makes none, that the first
paint bounds read after it does, and that a second read does not. The geometry tests gain
zero-width, combining, CJK, CR LF and negative letter spacing inputs and aligned paint bounds
with a trailing newline. Reading paint bounds before the first layout now asserts instead of
returning an inverted infinite rect.

Assisted-by: Claude Opus 5.5 (Medium)

First layout queried getSelectionRects once per cluster during line wrapping (~45% of cold layout time on large paragraphs). In default start alignment, each cluster's x is its left edge in LTR and its right edge in RTL while the vertical selection bounds are the span's font box, so TextSpan.extractClusters derives cluster advances directly from neighboring x coordinates and falls back to getSelectionRects only on tied x coordinates or non-monotonic negative spacing.

Perf: first_layout_short 500 -> 301 us (-39.8%, 95% CI [-40.0, +0.0], n=60/60) NOISE
Perf: first_layout_latin_long 6050 -> 3000 us (-50.4%, 95% CI [-54.1, -43.8], n=60/60) WIN
Perf: first_layout_large 1.45e+04 -> 6000 us (-58.6%, 95% CI [-59.7, -56.7], n=60/60) WIN
Perf: first_layout_multistyle 3600 -> 2000 us (-44.4%, 95% CI [-47.4, -44.4], n=60/60) WIN
Perf: first_layout_rtl_mixed 2401 -> 1699 us (-29.2%, 95% CI [-36.0, -29.2], n=60/60) WIN
Perf: first_layout_many_lines 6900 -> 3300 us (-52.2%, 95% CI [-52.9, -51.5], n=60/60) WIN
Assisted-by: Gemini 4 Argon (Medium)

AllCodeUnitFlags previously ran Intl.Segmenter for both grapheme and word boundaries eagerly during construction, even though layout only queries whitespace and line-break flags. Populate whitespace and V8 line-break flags eagerly and compute wordBreak and grapheme flags on the first matching hasFlag query.

Perf: first_layout_short 300 -> 201 us (-33.0%, 95% CI [-42.9, +0.0], n=60/60) NOISE
Perf: first_layout_latin_long 2850 -> 2250 us (-21.0%, 95% CI [-31.2, -12.7], n=60/60) WIN
Perf: first_layout_large 6000 -> 4500 us (-25.0%, 95% CI [-27.9, -23.7], n=60/60) WIN
Perf: first_layout_multistyle 1950 -> 1500 us (-23.1%, 95% CI [-25.0, -21.1], n=60/60) WIN
Perf: first_layout_rtl_mixed 1700 -> 1399 us (-17.7%, 95% CI [-23.5, -15.2], n=60/60) WIN
Perf: first_layout_many_lines 3300 -> 2600 us (-21.2%, 95% CI [-24.3, -18.8], n=60/60) WIN

Assisted-by: Gemini 4 Argon (Medium)

…ster bounds

Spans are in logical order and single-run spans emit clusters in ascending start order, so TextLayout.extractTextClusters only needs to sort allClusters when a multi-run BiDi span emits clusters out of logical order. Cache paragraph-wide start and end UTF-16 offsets on TextCluster so cluster insertion, mapping, and line wrapping avoid repeated span.start additions.

Perf: first_layout_short 201 -> 201 us (+0.0%, 95% CI [-33.1, +49.3], n=60/60) NOISE
Perf: first_layout_latin_long 2200 -> 2100 us (-4.5%, 95% CI [-8.7, +4.6], n=60/60) NOISE
Perf: first_layout_large 4400 -> 4200 us (-4.6%, 95% CI [-8.9, -2.3], n=60/60) WIN
Perf: first_layout_multistyle 1500 -> 1400 us (-6.7%, 95% CI [-6.7, +0.0], n=60/60) NOISE
Perf: first_layout_rtl_mixed 1400 -> 1400 us (+0.0%, 95% CI [+0.0, +7.1], n=60/60) NOISE
Perf: first_layout_many_lines 2450 -> 2500 us (+2.0%, 95% CI [-4.0, +4.2], n=60/60) NOISE

Assisted-by: Gemini 4 Argon (Medium)

Both callers of _clusterSelection guard empty text ranges before mapping, so assert at the helper boundary that the mapped cluster range is non-empty and within allClusters bounds.

Assisted-by: Gemini 4 Argon (Medium)

Canvas2D start alignment anchors the first visual cluster at 0.0 in both LTR and RTL, so only enter the single-run fast path when lastX == 0.0. Wrap three TextCluster comments to stay strictly within 100 columns.

Assisted-by: Gemini 3.8 Flash (High)

…, and didExceedMaxLines on re-layout

Normalize disjoint bidiLinePhysicalTextRange.intersect(span) slices to
TextRange.collapsed(bidiLineSpanTextRange.start) so a clusterless character
before a styled hard line break neither passes an inverted range to
_mapping.toClusterRange nor leaves (-1, -1) in TextBlock.physicalTextRange. Fall
back to span endpoint selection queries for block and trailing-space advances
when negative spacing inverts visual cluster order inside a TextSpan. Reset
paragraph.didExceedMaxLines alongside ellipsisClusters at the start of wrapText
so a wider re-layout after truncation clears maxLines overflow state.

Without these fixes, the three new tests in paragraph_relayout_test.dart fail
with ArgumentError 'TextRange [6:5) is out of paragraph text range',
Rect(0.0, -19.0, 28.3, 5.0) instead of Rect(0.0, -19.0, 21.1, 5.0), and
didExceedMaxLines == true instead of false.

Test: felt test --suite chrome-dart2js-webparagraph-ui test/webparagraph/paragraph_relayout_test.dart (47 tests)

Assisted-by: Gemini 4 Carbon (High)