fix(codegen): preserve hard-break spaces in height-auto text

[owjs3901]

A space the designer typed before an explicit line break is collapsed away by CSS, so a paragraph reads one way in Figma and another on screen.

Screen	Before	After
about-422-3376	5.51%	5.43%
about-422-3180	3.01%	2.93%
landing-833-3640	4.99%	4.78%
notice-422-6914	3.84%	3.74%

The other eleven screens are unchanged.

This continues an existing rule rather than inventing one

W8 established that painted text must equal characters - it stopped JSX inserting a space the design does not have, and stopped a design line break becoming a space. It then reported the remaining collapse as an explicit finding because no emission it had could avoid it. pre-wrap avoids it for a bounded case, so that case moves from reported to fixed and DEVUP_CODEGEN_TEXT_WHITESPACE_COLLAPSE stays for everything still unavoidable. One lossy fidelity impact is retired rather than left double-counted beside the diagnostic.

The condition is narrow on purpose: TEXT whose inline width Figma fixes while height grows (textAutoResize == HEIGHT), no positive maxLines, no list options, no tabs, and a space actually adjacent to a break. HUG text is excluded because pre-wrap counts trailing spaces in max-content width and would change intrinsic sizing; clamped text keeps its existing projection.

On the four goldens

Four plugin goldens gain six whiteSpace attributes. This is the opposite case to Korean keep-all, and the distinction is the whole argument.

With keep-all, matching Figma would have produced worse Korean - words chopped mid-break - so the compensation was deliberately kept and the pixel gain rejected. Here the design contains a space and CSS eats it, so preserving it makes the output match the source. The plugin drops that space; here the plugin is the one diverging from the design.

Each golden was reviewed against its collected fixture: 165 (node 1:18), 191 (107:32), 252 (284:18751/18753/18754), 253 (213:7494). Six attribute additions, no other output change, only those four checksums touched, corpus back to 268/268.

docs/about-landing-mobile-golden-review.md records the per-node review so a future reader can tell these two cases apart without re-deriving them - they look superficially identical and the answers are opposite.

Verification

Measured independently with binaries built from main and this branch, full 15-screen acquire and render on each side, repeated. My figures match the worker's to the decimal, and the baseline run fails correctly against all four tightened thresholds - confirming the two runs measure different code rather than reusing inputs.

New label derived-hard-break-whitespace states it preserves source whitespace without claiming intrinsic HUG sizing, glyph parity or identical wrapping. No viewport, breakpoint or node id in production code.

Gates: fmt 0, clippy 0, cargo test --workspace 1084 passed / 0 failed / 2 ignored, insta --check no drift, corpus 268/268.

Still open

A separate hero-export defect is quantified in the report but not fixed: it needs an isolated-fill export contract with crop and multi-paint validation first. Left documented rather than forced.