feat(layout): break lines within CJK runs and support word-break

[danperks]

What & why

Inline layout treated the text between breaking spaces as one atomic unit, so a line could only ever
be broken where the author wrote a space. That is fine for Latin prose and wrong for scripts that
don't use spaces: a Japanese or Chinese paragraph was a single unbreakable "word", so it never
wrapped, overflowed its container, and laid out one line tall where it should have been many.

This adds a UAX #14 subset covering the CJK case:

• a break is allowed between two ideographic characters (LB18/LB999),
• never before a closing bracket, sentence-ending mark, or non-starter such as a small kana or a
  prolonged sound mark (LB13/LB16/LB19) — those may not begin a line,
• never after an opening bracket (LB14) — it may not end a line.

Hangul is deliberately excluded: Korean is space-separated and UAX #14 only permits breaking between
syllables under an explicit line-break value, so treating it like Han would wrap Korean mid-word.

How it fits the existing line-filling loop

Rather than reworking the loop, a break opportunity inside a word is expressed by emitting several
InlineItem::Words from one source word, which the existing wrapping logic already handles. Because
an ideographic break is an opportunity and not a separator, the pieces are rejoined with nothing
between them — leads_space is set on the first piece only, and it drives both the width measured
during line filling and the rejoin, so a run that stays on one line paints byte-identically to
before. The same segmentation is applied in collect_inline_words, so min-content width agrees with
what can actually be placed on a line (grid, flex and table sizing depend on that).

word-break

The property was previously unparsed, and it has to be honoured here because normal, break-all
and keep-all select between three different sets of opportunities. The pair prohibitions apply to
all three — no value of word-break may strand a full stop at the start of a line.

keep-all suppresses the letter-to-letter opportunities but not the one after
U+3000 IDEOGRAPHIC SPACE: that character is a space rather than a typographic letter unit despite
sitting in the CJK punctuation block, and it is the only thing that lets keep-all text wrap at all.
WPT word-break-keep-all-005 covers exactly this.

Measured impact

Matched A/B over css/css-text (1,964 tests), same harness and same machine, only the code differing:

	tests
Baseline PASS	353 (18.0%)
With this change	456 (23.2%)
Newly passing	105
Newly failing	2
Net	+103

I also ran the same directory twice on the identical post-change build to establish a noise floor:
±1 test, and the one unstable test is in the shaping family below. So +103 is signal rather
than run-to-run variance.

The two newly-failing tests

shaping/shaping-024 and shaping/shaping-025 are Mongolian shaping tests that download
NotoSansMongolian-regular.woff2. They are flaky independently of this change: running
css/css-text/shaping three times on one unchanged build gives

shaping-023.html  FAIL, FAIL, PASS
shaping-025.html  FAIL, PASS, FAIL

shaping-024 fails in all three of those runs and passed once pre-change. Mongolian (U+1800–U+18AF)
is not in the ideographic set, and with default word-break this change is behaviour-identical for
non-CJK text, which neither test declares.

Breadth check, and a pre-existing bug this exposes

I also ran a matched A/B over css/css-writing-modes css/css-text-decor css/css-flexbox css/css-grid
(5,295 tests) because the change touches the shared inline path including vertical writing mode. Net
−3, all three in css-grid/grid-items/grid-item-block-axis-content-contribution-00{1,2,3}.

Those are worth explaining, because they are not a fault in this change. They style the item
font: 10px/1 Ahem and rely on 15 X glyphs measuring exactly 150px in a 150px box. The CSS
font shorthand is not implemented, so neither font-size nor font-family is applied and the
text is measured at the default 16px in the fallback face. Measured advance per character:

declaration	advance
font-family: Ahem; font-size: 10px	10.00px	correct
font: 10px Ahem	10.70px	wrong
font: 10px/1 Ahem	10.70px	wrong
font: 10px/1 "Ahem"	10.70px	wrong

At 10.70px the 15 glyphs need 160px and genuinely do not fit the 150px box. Previously break-all
was ignored, so the run could not wrap and stayed one line, coincidentally matching the reference.
Now that the opportunities exist the engine correctly wraps overflowing text — the new behaviour is
right and the wrong input is the font metrics. Respelling that one rule with longhands makes the item
lay out on a single line as the test intends.

I've left that out of this PR to keep it focused, but it looks high leverage: 6,573 of 49,011 files
under css/ (13.4%) use the font shorthand, most of them precisely to pin down Ahem metrics. Happy
to send it as a follow-up if you'd like it.

🤖 How this was built

• Authored primarily with an LLM (which: Claude Opus 5, driven via Cursor)

Checklist

• PR title is a Conventional Commit
• cargo test --workspace passes — adds linebreak unit tests (pair prohibitions, keep-all,
  break-all, the U+3000 case), ideographic_segments_are_placed_without_gaps covering the
  rejoin, a_run_that_exactly_fills_the_line_does_not_wrap, and
  word_break_parses_and_inherits
• cargo fmt --all + cargo clippy clean for this change (clippy reports pre-existing warnings
  in wurl, which this PR does not touch)
• WPT report shows no conformance regressions — +103 net on css/css-text; the 2 remaining
  shaping failures are demonstrated flakiness and the 3 css-grid ones are the pre-existing
  font shorthand gap, both analysed above

Made with Cursor