I've isolated the test with:
```diff diff --git a/dlls/user32/tests/win.c b/dlls/user32/tests/win.c index cbdd4a7b591a..9cd2bd62fbd9 100644 --- a/dlls/user32/tests/win.c +++ b/dlls/user32/tests/win.c @@ -14408,6 +14408,9 @@ START_TEST(win)
if (!RegisterWindowClasses()) assert(0);
+ test_initial_surface_colour(); + return; + hwndMain = CreateWindowExA(/*WS_EX_TOOLWINDOW*/ 0, "MainWindowClass", "Main window", WS_CAPTION | WS_SYSMENU | WS_MINIMIZEBOX | WS_MAXIMIZEBOX | WS_POPUP | WS_VISIBLE, ```
And ran it on the testbot and it seems to me that it fails quite often (https://testbot.winehq.org/JobDetails.pl?Key=161060, https://testbot.winehq.org/JobDetails.pl?Key=161061).
It makes me think that maybe the test suffers from the compositor more often than anticipated and that it is going to be failing spuriously already and should be dropped.
Adding a `Sleep(1000)` next to the `DwmFlush` call indeed seems to help (https://testbot.winehq.org/JobDetails.pl?Key=161067, https://testbot.winehq.org/JobDetails.pl?Key=161068) but it's probably still going to be timing dependent and I'm fine with the fix without a test.