feat(windows): per-window capture into a Flutter texture

[wenkaifan0720]

Implements the remaining v1 slice for the Windows backend, so both platforms now satisfy the full platform-interface contract.

Capture

Live capture via Windows.Graphics.Capture, delivered to Flutter as DXGI shared handles — pixels never round-trip through the CPU, matching what the macOS backend achieves with CVPixelBuffers.

Each frame is copied into one of two BGRA textures created with D3D11_RESOURCE_MISC_SHARED; the engine opens that handle directly. The pair rotates so the engine is never reading the surface being overwritten. The frame pool resizes itself when a source reflows, which is what keeps a live stream sharp after the user drags a window edge.

One-shot grabs (captureSingleFrame, captureSingleDisplayFrame) run on a worker with a 2s bound and encode through WIC, hopping back to the platform thread via a message-only-window task runner — a MethodResult cannot be completed off that thread.

Engagement events

This closes an honesty gap: the backend previously emitted only engaged and disengaged-on-release, so an externally moved or closed window silently desynced. SetWinEventHook now pushes the same three events the macOS backend does:

• alignmentBroken on external move/resize
• disengaged(windowClosed) on destroy
• disengaged(thirdApp | hostActivated) on focus change

Engage and conform mask their own moves for 1.5s, mirroring macOS's selfMutationDeadline. Minimize stays a poll, as it is on macOS. detectSeparateWindows is implemented rather than stubbed to empty.

Verification

Ran against real windows on Windows 11: 21 checks covering enumeration, both one-shot paths, stream lifecycle, geometry, and the event stream.

The decisive check rendered a live Texture and then re-captured the harness's own window — the result shows one window composited inside another, which proves the shared-handle path end to end rather than inferring it from a frame counter. That test also caught IsCursorCaptureEnabled and IsBorderRequired being properties rather than methods, so the ApiInformation::IsMethodPresent feature check was silently leaving cursor capture on.

Builds clean in Debug and Release under /W4 /WX; all four packages analyze with no issues.

Notes for review

• Display ids are HMONITOR values. Windows has no CGDirectDisplayID equivalent, so captureSingleDisplayFrame takes the monitor handle, and accepts '0' for the primary display.
• getContentMeasurement reports zero insets by design — DWM composites shadows outside the window, so there is no transparent margin to crop. A consumer treating zero as "not yet measured" rather than "nothing to trim" would behave differently here than on macOS.
• reconfigureStream reports rather than applies, since the capture pool resizes itself.
• Shared-file edits (root README.md, pubspec.yaml description) are prose only; the macOS backend is untouched.

🤖 Generated with Claude Code