Sigh. It is clear that my decomposition of the transformation matrices is incorrect; I get vastly incorrect rotation and sometimes incorrect scaling values. Banging my head against the desk didn't help either..
The problem seems to be that the matrices are assuming a left-handed coordination system, while my algorithms are for a right-handed one. So far, my research has failed to bring up anything useful.
Other than that I have a complete export of all RA2 components, and the minor scaling and rotation issues should be solvable by hand.