fix: preserve resolved text weights through a typography token

[owjs3901]

A text node whose segments share a typography token can still override the weight per run. Both text emitters suppressed fontWeight whenever a token was present, so alternating 400/700 runs came out uniform - and since missing bold changes glyph advances, a lost weight became a layout defect: one 36px line wrapped where the design does not wrap, and every section below inherited the displacement.

Measured on about against Figma's reference PNG, paired against a baseline built from main, identical theme hash 67d6de70e679 on both sides:

Width	Before	After	Design height	Height before	Height after
360	7.44%	7.46%	7240	7240	7240
992	6.90%	4.06%	5619	5656	5620
1920	4.19%	2.41%	4757	4794	4758

The 37px overshoot at both wider widths collapses to 1px. landing, popup, grid and keyframes are unchanged. Mobile's +0.02pp is the cost of correct bold runs and is left in the record rather than smoothed away.

Segment-sourced weights are attributed to styledTextSegments in the source map; explicit node weights keep their node-property mapping, so the provenance says which one supplied the number.

Korean keeps word-break: keep-all

The worker also implemented removing the Korean keep-all heuristic. Figma breaks Korean within words, so deleting it measurably narrows the gap - 7.46% to 7.25% at 360, 4.06% to 3.98% at 992.

It is not shipped, by the owner's decision. It buys 0.21pp on one screen by chopping every Korean word in every generated screen, and it breaks 38 attributes across 17 plugin byte-parity goldens. Figma's Korean line breaking is a limitation to compensate for, not a specification to reproduce - the same call the plugin makes deliberately in its own text renderer:

// Add wordBreak: keep-all for Korean text
if (hasKorean) {
  defaultProps.wordBreak = 'keep-all'
}

korean_characters_keep_words_whole locks it so a later fidelity pass cannot quietly reverse it, with the reasoning recorded on segments_contain_korean. Keeping it is also what holds 16 of the 17 goldens at parity: one golden changes here, for a segment that explicitly carries weight 400.

On the numbers

thresholds.json is deliberately unchanged. The harness does not reproduce absolute figures across sessions - the same commit and binary produced about theme d3137cabf5a5/11.24% in one session and 67d6de70e679/7.44% in another, and isolating each group into a fresh acquisition process does not fix it. Only the paired delta is trustworthy, so lowering an entry would assert a reproducibility that does not exist yet. Harness determinism is tracked separately; docs/about-vertical-geometry.md records the evidence.

Both the worker and I measured this independently and reached identical figures to the decimal.

Gates: fmt 0, clippy --locked --workspace --all-targets --all-features -D warnings 0, cargo test --workspace 1059 passed / 0 failed / 2 ignored, insta --check no drift, stdio_smoke 2 passed.