since the docs are lacking in that regard, your options are to read the source and/or experiment and see what happens.
you might find some answers in relevant footnotes, which hopefully point to books. a lot of these algorithms are implementations from original publications (books, papers, lectures).
it should be possible to use dissimilar cameras and picture sizes. at some point, matrices will be calculated such that both camera views are scaled (and undistorted and rectified) so they’re comparable… i.e. suitable for block matching.