Initially thought this was a bug in my script, but I've managed to reconstruct the problem in a simple blend file and am pretty sure this is a python/mathutils/matrix bug. 2.55 beta windows 32 from blender.org.

Open the blend, select the empty, click run script until the empty disappears (some times it only takes on press of run script, sometimes its 10 or 20 before it goes).

The empty is shrinkwrap constrained to the plane with a shrinkwrap. The very short python code in the top window displays the location and the translation part of the world matrix of the active object (the empty). It copies the translation part across to the location (effectively applying the SR without deleting the modifier). The code then prints the location and the translation part of the matrix. NOTE they are now the same in the terminal and are both real numbers. Look in the 3D view and the empty has disappeared and its location in the transform panel displays (0,0,0). g-key grab the empty and note the location transform fields turn to '-1.#IO'.

Its the same effect as if I was to set the location with a vector full of 'NaN's but python has explicitly told me in the terminal that my matrix does not have 'NaN's in it and that immediately after setting the location, its value is correct, not zero and not a NaN.

If you run the script several times the print-outs in the terminal will eventually change to NaNs.

Many Thanks,
Josh