textarea: expose cached visual row count

[TomiXRM]

Problem

A host that owns a textarea's preferred height cannot currently query its cached visual content height in rows. rows and auto_grow control layout constraints, not the total number of wrapped content rows. Using auto-grow as a minimum also prevents a surrounding flex layout from shrinking a composer in a short window.

Change

Expose TextareaState::visual_row_count() -> usize as an O(1) read of the existing display-map cache. It counts explicit and soft-wrapped content rows independently of minimum rows, auto-grow limits and viewport height. It does not shape or copy text, allocate, create a second cache, or force layout.

Document the cache contract: it may be zero before first layout; laid-out empty text has one row; width/font changes become visible after layout. Hosts can keep rows(1), constrain the outer preferred height, and consume the refreshed count on a subsequent render.

Coverage exercises pre-layout behavior, fixed one-row textareas, preferred 3–12-row sizing, constrained flex shrink, internal scrolling and deletion, Unicode wrapping, width changes, and disabling soft wrap.

Verification

• cargo test -p gpui-base --lib: 1,375 passed.
• cargo test -p gpui-kit --features test-support --test integration --test ui: 395 passed.
• cargo clippy -p gpui-base -p gpui-kit --lib --tests --features gpui-kit/test-support: passed; only the existing block 0.1.6 future-incompatibility notice.
• cargo fmt --all -- --check: passed.
• Throwaway native macOS Metal smoke program (removed afterward): 3/6/20 content rows, a 120px constrained viewport, internal scroll offsets 0/-18/-298px, and real Unicode text wrapping from 4 to 10 rows when narrowed from 480px to 200px. Captured using the native text system and renderer, not the test renderer.

Downstream motivation: Kagi's official-dependency migration. This API removes the last cached-row accessor requirement without another wrapping implementation or a dependency patch.

[huacnlee]

Thanks for the thorough write-up and tests. The use case is valid, but we'd rather not add a new accessor for it, so I'm closing this PR.

DisplayMap is already public, and display_map().wrap_row_count() is the documented way to size to content (see "Sizing to content" in website/component/editor.md). The only gap is that display_map() (and prewrap) live in the EditorMode-only impl block, so TextareaState can't reach them.

The fix we'd accept: move display_map() and prewrap into the impl shared by all modes, so a textarea host reads text.read(cx).display_map().wrap_row_count() exactly as an editor host does. The cache notes from your doc comment (zero before the first layout, refreshed by layout, prewrap for the first frame) belong on those methods.

You're welcome to open a new PR along those lines. Your integration tests (preferred height with flex shrink, internal scrolling, Unicode wrapping and resize) would carry over directly on top of display_map().wrap_row_count(). Please also keep the docs change minimal by pointing textarea.md at the same pattern as editor.md, and sync it to website/zh-CN.

[TomiXRM]

ご指定の方針に合わせた新しい PR #3564 を作成しました。既存の display_map() / prewrap() を共通 impl へ移し、専用 getter は追加していません。キャッシュの説明、Textarea の初回レイアウト・preferred height/flex shrink・内部スクロール・Unicode/resize の検証、英語/中国語の最小限のドキュメントを含めています。macOS native Metal の描画証跡と実行した検証コマンドも新しい PR に記載しました。