I'm debugging a reproducible X focus problem with gtk2 builds of Mozilla
variants using the Ion window manager. The problem details are in Mozilla
bug #230097 (https://bugzilla.mozilla.org/show_bug.cgi?id=230097).

Debug traces (see comment #12 in the above bug report) seem to indicate
that the problem might lie in the gtk+ library. My limited understanding
of the Mozilla code is that it's relying on the gtk+ focus events to keep
track of which of its windows has the keyboard focus. The problem seems to
be that these events are not dispatched under certain conditions although
the focus window has changed.  These conditions depend on the mouse
pointer, and are connected to X focus events with detail NotifyPointer.

Tuomo Valkonen, the author of the Ion window manager, says in the
Mozilla bug report that the NotifyPointer X focus events don't contain
actual focus information but Mozilla seems to be taking them for normal
focus events. This seems to actually take place in gtk+: see below.

Is this the intended behaviour or a bug in gtk+ ?

The gtk+ source code (version 2.4.13), file gdk/x11/gdkevents-x11.c,
looks to me as the focus_in_event and focus_out_event events are
dispatched from gdk_event_translate() based on the received X focus
events, when the HAS_FOCUS state of the window changes. The HAS_FOCUS
macro is defined as

#define HAS_FOCUS(toplevel)                           \
  ((toplevel)->has_focus || (toplevel)->has_pointer_focus)

Now gdk_event_translate() sets/clears 'has_focus' by these X focus event details:
            case NotifyAncestor:
            case NotifyNonlinear:
            case NotifyVirtual:
            case NotifyNonlinearVirtual:
and 'has_pointer_focus' just by this:
            case NotifyPointer:

In the problematic case the NotifyPointer X focus events are unpaired,
ie. the window gets a NotifyPointer FocusIn event, but not a NotifyPointer
FocusOut one. This leads to the HAS_FOCUS state remaining true after the
WM has set the focus to another window (has_focus is false but has_pointer is
still true). Thus a focus_out_event is not dispatched. When the WM
sets the focus back the HAS_FOCUS still remains true (only has_focus
changes), and no focus_in_event is dispatched.

I hope I'm making myself clear; if you need traces or anything else,
I'll be happy to provide them.