Closes #5393.

Problem

Dragging an image's resize handle and releasing the mouse button outside the MarkText window discarded the resize. The image kept the new size on screen, but the document was never marked modified, Save wrote the line back as ![](pic.png) with no width, and reopening the file showed the original size.

Root cause

A pointer dragged past the window's edge keeps driving the resize — the browser holds the pointer while a button is down — but Chromium hit tests those out-of-viewport coordinates to <html>. <html> is not inside <body>, so the event's bubble path is html → document → window and skips <body> entirely, which is where ImageResizeBar attached the drag's mousemove and mouseup listeners. _mouseUp therefore never ran, so updateImage was never called. Chromium then delivers a click on the press/release common ancestor (again <html>), which the bar hides on, and hide() clears the pending _width (#5288) — the width the drag had written straight onto the <img> element stayed on screen with nothing behind it.

Verified with OS-level mouse input (CGEvents) against a real Chromium window:

BODY:mousedown target=#handle x=110 y=125     <- body listener fires
BODY:mousemove target=BODY   x=300 y=125      <- inside the window: fires
DOC :mousemove target=HTML   x=780 y=125      <- outside: body listener silent
DOC :mouseup   target=HTML   x=780 y=125      <- _mouseUp never runs
DOC :click     target=HTML   x=780 y=125      <- hide() -> _width = null

Fix

Attach the drag's mousemove/mouseup to document instead of document.body, so the listeners sit above the <html> target. tableDragBar, paragraphFrontButton and TableRectSelection already attach their drag listeners this way — imageResizeBar was the only one on document.body. The mousedown listener stays on document.body; the handles always live there.

Tests

• packages/muya/src/ui/imageResizeBar/__tests__/imageResizeBar.spec.ts — a drag whose moves and release are dispatched on document.documentElement (followed by the click Chromium delivers on <html>) must still commit the width, and must leave no drag listeners behind. The first test fails on develop with updateImage never called.
• packages/muya/e2e/tests/drag/image-resize-outside-window-5393.spec.ts — a real Chromium drag released past the viewport's right edge. It asserts the release target really was HTML before asserting the markdown gained a width, so a clamped or shortened drag cannot leave the regression uncovered.
• Two existing assertions in the unit spec keyed on e.target === document.body were narrowed to e.event === 'mousemove'; with the drag listeners on document they would otherwise have passed vacuously.

Also verified end to end in the built app with real OS-level mouse input, dragging a handle out through the window edge and releasing there:

🤖 Generated with Claude Code