I have to look on the velocity values again. I did not verify if the
values are in their boundaries now. But I did not see the
backward-driving any more. It was 100% reproducible when the movement
started in a position where the robot would have to turn 180° in order
to find free space. The same situation now is mastered by the local
planner superbly.

So it really seems that the CostMap timeout broke something in the
overall behaviour.