I (along with several people who came to my aid) tried for almost two years to get arXiv to solve this problem sensibly. I've now decided to go public, and the reactions have been overwhelmingly favourable. While things were going on only behind the scenes, I had some password-protected web pages so that people could read up on the details; I have now made them public.
For non-astronomers: MNRAS, Monthly Notices of the Royal Astronomical Society, is one of the most respected journals in the fields of astronomy, astrophysics, and cosmology. That a paper which has appeared there is considered unacceptable for the astro-ph category at arXiv is patently absurd.
The problem is that as long as arXiv claims the right a) to reclassify any paper b) without even as much as an automatic email to inform the author and c) with no possibility to completely delete the submission, that means that I would live with the risk of having the paper appear in the wrong category, where those interested would not find it. (Sure, if one knows that it exists, one could find it on arXiv, but if one knows that one exists, one can find it elsewhere; the only reason for me to put something on arXiv is visibility: I have the stamp of approval from the journal, I host a copy on my own webserver, and ADS has links to the version I host.) An appeal is theoretically possible, but my experience is that that is a farce, not only because arXiv does not follow its own documented appeal procedure. (Even if successful, it is not clear to me that it would show up in the list of "new" abstracts, which is the whole point.) In the case of the paper mentioned above, I was exceptionally allowed to delete it, then was accused of violating the rules because I did so. (Note that technically unsubmission and deletion were possible, but since authors are not informed of a reclassification, most won't have a chance.)
I'll mention the names of those I dealt with (in addition to anonymous emails, some of which are probably from humans and some automatically generated), because they are all guilty of at least one of the following: not following arXiv's own documented procedure in the case of an appeal, arrogance, unfriendliness, unhelpfulness, at least distributing misleading information (probably lying): Steinn Sigurdsson (still Scientific Director of arXiv), Eleonora Presani (then Executive Director of arXiv; apparently no-one is now), Licia Verde (then Chair of the arXiv advisory board), and Robert Seiringer (then Chair of the physics committee). I'll write up my ordeal in more detail elsewhere. In the meantime, people who are genuinely interested in getting arXiv to serve the community, rather than vice versa, especially if they (believe that they) can actually influence arXiv's behaviour, directly or indirectly, can contact me for further details.
I'm interested in hearing from people with similar experiences.
Several people (all professors of cosmology at major research universities) tried to convince arXiv to let my paper appear in astro-ph (as is the case with all other papers which I have submitted to arXiv), but to no avail. As such, I've decided that it is not worth the trouble to bother with arXiv until they reform their procedures and institute a more transparent workflow.
I've had better experience with the PhilSci ARCHIVE, where I have one paper already, but most of my papers are not about the philosophy of science (and even in such cases it would make sense for them to be on arXiv as well).
Anyone who believes that everything worth reading in their field is at arXiv gets what they deserve (or, better, doesn't get what they don't deserve). On top of that, those who read only stuff at arXiv are not in a position to know whether there is anything worth reading which is not at arXiv. If you think that arXiv's policy here is wrong, feel free to complain about it. My experience is that many share my views but most are afraid to let that be known for fear of getting banned by arXiv, the modern-day equivalent of excommunication. Fortunately there are still good journals which make choices based only on quality, are not prohibitively expensive (I have never paid any sort of publication costs, article-processing fees, page charges, or whatever, neither directly nor indirectly, and always publish with journals which allow at least the author's accepted manuscript to be distributed freely, though perhaps not at arXiv). arXiv would be a good idea if it actually served the community better, but any archive which does not allow papers from the leading journals in the field into the appropriate category is seriously broken and it is a shame that the community is not able to do anything about it.
For arXiv-overlay journals, the result is that they have to accept arXiv's decision as to what is appropriate. That means that they are essentially limited to papers put on arXiv before submission to their journal. While that might be the norm, especially for arXiv-overlay journals, there are a significant number of people who, for a variety of good reasons, prefer to put stuff on arXiv only after acceptance. While that is technically still possible with some arXiv-overlay journals, most probably don't want to take the risk of having a paper accepted by an arXiv-overlay journal but not having it appear on arXiv. In fact, while the journal can accept the paper before it appears on arXiv, it cannot publish it, since it relies on arXiv for publication. Thus, such journals—arXiv-overlay journals which allow the possibility to submit to the journal before submitting to arXiv, in itself a good thing—cannot really call themselves journals since one of the main tasks of a journal is to publish papers.