Modern console issue: initial screen buffer can never be freed

[rprichard]

In Win8.1 and Win10 (and presumably Win8 as well), it's impossible to free the initial screen buffer. It is possible to enter a state where the initial buffer can't be accessed via any means, but I can still see it taking up memory via the Task Manager.

Test 1:

// Tie the child to a new buffer, then kill the parent.  There is no way to
// reference the initial screen buffer, but it's still taking up memory.
int main() {
    Worker p;
    p.getStdout().setFirstChar('A');
    p.getStdout().trySetScreenBufferSize({ 30000, 1000 });
    p.newBuffer(false, 'B').activate();
    auto c = p.child();
    p.exit();
    Sleep(60000);
}

Test 2:

// This test is identical to Test 1, except that the parent reactivates the
// initial screen buffer just before it exits.  I would have expected buffer A
// to be freed after the parent died, activating the remaining buffer B,
// because SetConsoleActiveScreenBuffer does not normally prevent a buffer from
// being freed.  The initial buffer is never freed, though.
int main() {
    Worker p;
    p.getStdout().setFirstChar('A');
    p.getStdout().trySetScreenBufferSize({ 30000, 1000 });
    p.newBuffer(false, 'B').activate();
    auto c = p.child();
    p.getStdout().activate();
    p.exit();
    Sleep(60000);
}

This behavior does not apply to Windows 7. e.g.:

int main() {
    Worker w;
    w.getStdout().setFirstChar('A');
    w.getStdout().trySetScreenBufferSize({ 30000, 1000 });
    auto b = w.newBuffer(false, 'B');
    w.getStdout().close();                 // -- L1 --
    w.getStderr().close();                 // -- L2 --
    CHECK_EQ(w.openConout().firstChar(), 'B');
    Sleep(30000);
}

If I comment out lines L1 and L2 above, then the CHECK fails, and the RAM usage of conhost.exe increases to ~60MB on Win7 32-bit. Otherwise, it's only about 1.5MB.