This turns out not to be Windows-specific. There are at least two defects in the test.

1. The test does not wait for the processKiller callback to return, so it may fail after the test function has already returned (as observed here).
   

   go/src/os/os_test.go

   Lines 2308 to 2311 in 0c5afc4

   	go func() {
   	time.Sleep(100 * time.Millisecond)
   	processKiller(cmd.Process)
   	}()

2. The callback invokes t.Fatalf from a non-main goroutine — which may invoke runtime.Goexit — but does not propagate the Goexit back to the caller. It should probably use t.Errorf instead.
   

   go/src/os/os_test.go

   Lines 2328 to 2333 in 0c5afc4

   	testKillProcess(t, func(p *Process) {
   	err := p.Kill()
   	if err != nil {
   	t.Fatalf("Failed to kill test process: %v", err)
   	}
   	})

Change https://golang.org/cl/284153 mentions this issue: os: invoke processKiller synchronously in testKillProcess