.mp files cater for different zoom levels and perhaps oddly, different levels rely on different lats and lons data; if you examine an mp file you will notice slight changes in the coords of a particular node depending on your zoom level (data1,data2,data3). This fine tuning is ignored by osm. I’m currently working on an mp2osm conversion gui.
I can’t really do the sort of code examination you describe, because all the objects are defined on Data0 with an End Level (generally 2) so there are no Data1s, Data2s, etc. to compare. Also, the blank space is far too large to be accounted for by the increased generalization of the higher levels, and the map routes correctly across the blank space.
Since the map renders correctly when compiled with an earlier version of mkgmap, I suspect that there may be a problem with the way later versions deal with tile edges. Could be something about the TDB or the base map IMG.