Well, I don't know the exact cause of the differences, but specifying the observer elevation and offset does let me control the outputs and I can reproduce similar patterns to both what I get in 9.3 and 10.3. My guess is it has something to do with the way they are doing bilinear interpolation to get the source elevation for the default (either that or in the application of the z factor to convert from ft to m). I couldn't test it directly, because ArcGIS 9.3 kept crashing when I tried to specify the observer elevation and off-set and I've run out of time to troubleshoot it.
My take-home from this is to run a sensitivity analysis so that the results can account for some uncertainty in observer height and surface height.