Yeah, I think (but am very happy to be corrected) that Germanic mainly crops up in introductory handbooks; so it's not being used actually to reconstruct anything anew, but simply to allow readers to compare the various families alongside standardly accepted reconstructions (which will not necessarily have been reconstructed from the "data-sets" given). And they can do their own kind of proto-reconstruction with the given forms just to see how the reconstructions came about, which is easier for introductory purposes than them having always to use the most valuable forms from languages they know little about. So whilst there may be no objective/scientific reason why modern German is chosen, rather than middle Aryan or whatever (which, in fact, as Lyceum says, would undoubtedly be a better choice when doing original reconstruction), I think the cultural reason is compelling enough, that lots of readers will be familiar with modern German but not middle Aryan -- and then giving German, alongside English, allows the reader to compare another modern Germanic tongue, since English is going to be there anyway (if only to serve as a translation), and allows them to compare and contrast the effects, for example, of Grimm's Law in the two languages (which, of course, has nothing to do with reconstructing PIE, but is interesting for its own sake).
So, yes, it would be a methodological fault if those tongues were used in the reconstruction of PIE, but when just introducing students to the received wisdom, it seems quite useful for German to be included.
Also, Apeiron: it's been a while since I've read anything relating to PIE, so I may be a bit shaky here, but you say that one should assemble proto-Germanic, proto-Indo-Iranian etc. and use those forms to shed light on PIE, but I'm wondering how useful that would be, when we know, for example, that proto-Romance is markedly different from the form
of Latin that the Romance languages did actually evolve from (in other words, one doesn't get a very accurate picture of the ancestor language because its descendants lost so much information)? So if you forgo the intermediate step, you're likely to get a more accurate result (since you won't be reconstructing from hypothetical forms that are bound to contain errors), right? Or not?