feat(text): Notepad word wrap + undo/redo + the wrapText host op (gpui LineWrapper)

[doodlewind]

Word wrap, complete: the app-layer layout model and its platform half — the same division of labor Zed uses.

The pattern (settled against Zed's source)

Zed's editor does NOT hand paragraphs to a toolkit text view. gpui provides fonts, per-char advances and LineWrapper (greedy width-fitting with break candidates); the editor's WrapMap — application layer — calls it and owns the wrapped coordinate space, caret movement and hit testing on top. PocketJS now has the identical split:

half	Zed	PocketJS
"where does this line break at width W"	gpui LineWrapper	wrapText op (spec 43)
wrapped coordinate space + caret/selection/hit	editor WrapMap	notepad.ts visual segments

The wrapText op (spec op 43)

wrapText(str, fontSlot, maxW) → u32[] — ascending UTF-16 break columns, empty = the line fits. Optional like every post-16 op; apps without it fall back to the same greedy rules over measureText.

• engine/core: greedy word wrap (break before the overflowing word, trailing spaces hang, over-wide words char-split) computed over whatever provider measure_run resolves to — atlas advances on baked hosts, the native measurer for native-text apps. Break positions always derive from the metrics that size and paint the text.
• gpui backend: native_wrap — gpui's own LineWrapper, resolved through the same TextConfig as native_measure and the painter; hosts/macos installs it next to the measurer under --native-text. Measurement, wrap and paint stay one provider.
• wasm + symbian: ui_wrap_text / ui_wrap_text_ptr staging exports; wasm-ops.js feature-detects them.
• pocket-ui-surface: the QuickJS op + set_text_wrap.

The app layer (desk98 Notepad)

The Doc keeps logical lines; wrapDoc() projects them onto visual segments {row, from, to}, and render, clicks, drag-select, visual Up/Down/Home/End (with wrap affinity at soft boundaries), selection highlight, scrolling and the IME caret all read that one segment list — reflow on window resize falls out of the reactive width. Edit > Word Wrap is a live checked toggle. desk98 consumes the op when present and keeps the JS greedy as the fallback.

Undo/redo (⌘Z / ⌘⇧Z)

Docs are immutable values, so history is plain snapshot references — O(1) per step (notepad.ts History, pure). Coalescing: typing runs and erase runs collapse into one undo unit each; Enter/paste/cut/Time-Date/New stand alone; a caret move between keystrokes breaks the group (continuity is checked against the doc the last edit produced). Undo restores caret and selection. One applyEdit door for every mutation (no-op edits record nothing); Undo/Redo rows join the Edit + context menus with live disabled states.

Verified

• tests/wrap-op.test.ts: the op column-for-column equal to the JS fallback over desk98's committed W95FA atlas (275 assertions — both reduce to the same additive advances).
• Core unit tests: greedy rules + native-wrapper override (122 core tests green).
• 4 pure wrap-model test groups (greedy/hanging/char-split, caret↔row mapping with affinity, visual movement, selection∩segment); the desk-companion sim journey extended (a two-keystroke run undoes with ONE ⌘Z, replays with ⌘⇧Z); bun run test 12/12; pocket check clean; gpui + macos clippy/fmt clean; bun run contract green (spec.rs regenerated).
• Sim renders through the op path are pixel-identical to the JS path (narrow-window reflow, cross-row selection, the ✓ menu item).

Follow-ups (not in scope): migrating apps/note/apps/im's own wrap math onto the op; richer break rules (UAX #14 CJK classes) — they'd land in ONE place now, core + LineWrapper already handle the western rules identically.

🤖 Generated with Claude Code