The reasons are not complicated: If I use any of the perl+XML machinery, it wants me either to let it read the whole thing and build a structure in memory, or go to a callback interface.

Since we're typically reading very large datasets, and typically looking at the vast majority of it, preloading it into a data structure would be impractical not to say stupid. Thus we'd be forced to use parser callbacks of one kind or another, which is sufficiently non-idiomatic and awkward that I'd rather just live in regexp-land.