[Copilot]

Bug Fix

What was the bug?

In mixed-orientation multi-monitor layouts, Fly Away computed end positions relative to only the source monitor’s work area.
That could place windows “off” one monitor but still visible on another (e.g., lower half of a vertical secondary display).

How did you fix it?

• Target-space correction (monitor-local ➜ virtual desktop)

  • Kept the existing direction heuristic (left/right, up/down) based on the window’s position within its current monitor.
  • Switched final offscreen target coordinates to the virtual desktop bounds so endpoints are outside all visible monitors.

• Interop support

  • Added missing GetSystemMetrics constants in NativeMethods:
    • SM_XVIRTUALSCREEN, SM_YVIRTUALSCREEN
    • SM_CXVIRTUALSCREEN, SM_CYVIRTUALSCREEN

• WindowTracker change

  • Added GetVirtualScreenBounds() and used it in ComputeFlyAwayTarget(...) when calculating final targetLeft/targetTop.

int targetLeft = moveLeft
    ? virtualScreenBounds.Left - width - OffscreenMargin
    : virtualScreenBounds.Right + OffscreenMargin;

int targetTop = moveUp
    ? virtualScreenBounds.Top - height - OffscreenMargin
    : virtualScreenBounds.Bottom + OffscreenMargin;

[shanselman]

Triage note: this is plausible, but it needs human/manual validation because the bug is specifically mixed-orientation multi-monitor Fly Away behavior. I would not take this automatically until someone verifies at least: horizontal+vertical monitor layout, windows fully leave the virtual desktop in each direction, restore still returns windows to original positions, and single-monitor behavior is unchanged.

[shanselman]

The branch is now current with main and hardened in f1806d5.

Production behavior now:

• enumerates the live topology with EnumDisplayMonitors once per Fly Away operation
• has no fixed monitor count, resolution, origin, orientation, or rectangular-layout assumption
• evaluates safe exits at every relevant monitor edge and chooses the shortest one-axis path
• requires a conservative 64px clearance from every monitor, not just the source display
• preserves window size and rejects overflow-clamped candidates

Validation:

• solution build
• complete interop harness at 10,000 iterations
• deterministic examples for portrait, negative-origin, staggered, L-shaped, oversized, and three-monitor dead-space layouts
• 2,000 broad randomized layouts (1-16 monitors)
• 300-layout independent pixel-scan optimality oracle
• independent adversarial audit with exhaustive/randomized counterexample testing
• x64 NativeAOT publish

No blocking geometry findings remain. This should stay draft until #74 is merged and a human verifies mixed-DPI/mixed-orientation hardware, hot-plug behavior, restore accuracy, and a Task Manager kill during Fly Away.