Yes the whole thing is a horrible mess, and quaternions (which have the advantage of having a uniform representation) are too much for normal people (well, me, anyway) to get their heads around in normal use. (Everything is presumably handled as quaternions internally.)
The problem which Martin is expressing is that rotation is not commutative or associative: AB != BA, A(BC) != (AB)C -- so XYZ rotation (which literally is performed in that order) is simply incompatible with YXZ.
I have a problem with the labels "HPB" because I sometimes have to think for a second to figure out which axis I'm rotating around, but I do not have a problem with the assumed order of rotation. I'd be happier with H(Y), P(X), B(Z) or something.