This recipe is not exactly reliable.  It usually takes up to 20 seconds
of mouse movement to trigger the issue here.  Some scrolling helps.

I realise my recipe puts a lot of pressure on the display engine.  When
reducing the timer's REPEAT value, the issue does not vanish but is
harder to reproduce.  My real use-case, which provokes this issue after
a few minutes, is a process filter which updates the buffer of a window
exactly once a second.

When applying this recipe to 24.3 on the same system using the same
libraries for building the binary, there are some glitches.  Most
notably the mouse face of mode-line-buffer-identification is not reset
reliably.  Yet 24.3.91 is a lot worse than 24.3.  IMO this is a
regression.

    In GNU Emacs 24.3.91.3 (x86_64-unknown-linux-gnu, GTK+ Version 2.24.10)
    Repository revision: 117154 eliz@gnu.org-20140525141846-w6719v7ajcd3tbkk
    Windowing system distributor `The X.Org Foundation', version 11.0.11204000