It is not possible to reliable simulate key-events when running in Xvfb, whereas
a "native" X-Server does not cause any problems.
Some key-events are received by the client-application twice, others
get completly lost.

I first thought this is a specific problem of the library I used for
this task (java.awt.Robot class in java) and wrote my own library in
plain C using XTest which is heavily based on the code of Xvkbd-2.7 and
it shows exactly the same sympthoms. If I insert a wait of 200ms after
each key everything works like expected with both libraries (100ms do
not work).

For my testing I used Xvfb delivered with Xorg-6.8.2, running with the
following flags:
Xvfb :1 -screen :1 500x1000x24 -fbdir /tmp/

I attached a a test-case which is derivated from xkdbd-2.7 which shows the
problem (have a look at main() of CRobot.c) and a simple bash-script to compile
the test-case.
Note: Although the Xtest-code is really ugly hacked withought any deep knowledge
whats going on behind the scenes (thats why I unmap Xkdbd's main-window
immediatly after creating it instead of not mapping it at all) I think its a bug
in Xvfb since I had absolutly the same problems when using java.awt.Robot from
Java which does only provide low-level keyPressed/keyReleased methods.
Note that this problem sometimes shows up and sometimes not.

I already tried to find it myself, however I think I lack too much x knowledge :-(

Good luck with fixing it, lg Clemens