I don't think there should be a "non-mainstream" section, the Q/A site itself is open enough to allow questions/answers about any topic, with open discussion, so long as nobody is censored, and sincerity is required. It is very easy to tell a sincere mistaken person from a troll.
Just say "no repeats", then crackpot material can appear once somewhere on the Q/A site, and any further reference to this idea can be simply through a link with a title that is revealing of the basic idea, so that the author does not feel censored, but people don't have to waste their time. If original material is incorrect, it will be downvoted, you need to trust the community. If it is crackpot material, it is nearly always unoriginal, and it can get a quick refutation. If it is original, even if it is crackpot material, it can be valuable. Mistaken orginal work is often interesting regardless of correctness. The easiest reliable way to repel dishonesty without repelling new ideas which are plausible and supported by evidence is to vote based on originality. Unoriginal material is easy to spot, and original material is always valuable, even when it is dead wrong. So just require originality for contributions, beyond no copy-paste.
The distinction between "flaky" and "reliable" is not so clear-cut, but the distinction between "original" and "rehash" is very clear. Original material is difficult to produce, it will be limited in size, and if it is wrong, the debunking can happen at one place only. The same goes for mainstream questions. Original insights are always more valuable, because the best assumption on the internet is that the readers already know everything old.
There are scaling issues with a site which has no evaluation of originality. There are lots of people who will want to rehash the same three topics, even if these are nice topics, like the standard model, or chiral anomalies, this does not usually add value, unless there is originality in the exposition. What adds most value is a discussion which cannot be found anywhere else. This is why scientific journals ask for original things, but not too original, or else it gets rejected, right or wrong.
The point of the internet is that it allows original work to be evaluated openly. You need to trust open evaluation to naturally downvote the crackpot material without any special dispensation or committee. If you find a crackpot thing which is highly upvoted, for instance, something regarding cold-fusion, it is likely an original thing which has some merit, and this is when you can see that the mainstream understanding needs revision.
There are lots of fringe ideas which might have merit. For example, in high energy physics, there are occasional "tetraquark" meson ideas (two meson bound states). A lot of these can be dismissed with simple estimates, but it is not completely clear that the sigma(660) or the f0(980) aren't tetraquarks, nobody knows the proper description. But tetraquarks have a bad reputation for a reason--- a lot of tetraquark putative resonances are really clearly just string excitations of QCD (this was pointed out to me by Henry Tye). A higher string excitation of a meson can easily look like a quark-anti-quark pair was inserted in the middle without changing the quantum numbers.
Another fringe topic with perhaps some interesting physics is ball-lightning. This has no model, and it is associated with UFOs and paranormal nonsense. The criterion which excludes most paranormal nonsense is physicist positivism, and since it is standard, it automatically comes for the ride in a physics website.