Fix stuck pan when middle-button release lands on toolbar overlay

[mucow24]

Summary

Middle-drag pan was sticking to the cursor (with no buttons held) when the user released the middle mouse button while hovering over the floating top toolbar — pan only stopped on the next click anywhere.

Root cause

Two compounding issues:

• Dual-path pan architecture: Pixi pointermove drove pan when the cursor was inside the canvas; a window mousemove listener drove it when outside. The toolbar overlays the canvas, so neither handler could see the release once the cursor was over a toolbar element.
• No button validation: Neither move handler checked that any mouse button was actually held, so once panDragRef was set, any future move kept panning.

What changed

• Collapsed pan to a single window-level pointer-event listener (pointermove / pointerup / pointercancel).
• Switched from mouse events to pointer events because Chromium fires only pointer events for CDP-driven middle-button gestures (verified via in-browser console diagnostic).
• Added a defensive event.buttons === 0 recovery so a missed pointerup self-heals on the next move.
• Removed the dual-path Pixi pan handler, the isClientPointInsideViewport early-return, the isPanDragging state, and the handleScenePointerMove plumbing through stageHandlers and CanvasStage — all dead after the collapse.

Net diff: -36 lines across 7 files.

TDD

Added a red-first jsdom test: while pan is active, dispatch a window pointermove with buttons: 0 and assert the pan was cleared. Verified failing before the fix, passing after.

Existing pan tests adapted to dispatch PointerEvent with buttons set (a PointerEvent polyfill was added in src/test/setup.ts since jsdom doesn't ship one). The synthetic test API in canvasTestApi.ts now tracks pressedButton through middle-button-only startPanDrag paths so subsequent pointermove dispatches report buttons=4 instead of 0.

Test plan

• npm run preflight (lint + typecheck + unit + e2e)
• New jsdom regression test passes
• Manual: middle-drag pan from canvas, slide cursor onto top toolbar, release middle button there, move cursor back over canvas — pan must not follow
• Sanity: spacebar+left-drag pan still works; pan-tool (Hand) drag still works; normal middle-drag release on the canvas still ends the pan

No Playwright e2e for the regression itself: headless Chromium on Linux/macOS doesn't reproduce the Windows middle-button trigger that surfaces the bug, so a passing e2e would be misleading coverage.

🤖 Generated with Claude Code