What processors do/do not support PAE?

You should be pretty safe assuming PAE for any Pentium II or Athlon or newer, although some Pentium M's (marketed as Centrino)--namely, those with a 400 MHz bus--do not support PAE.

http://en.wikipedia.org/wiki/Physical_Address_Extension

PAE is provided by Intel Pentium Pro (and above) CPUs - including all later Pentium-series processors except the 400 MHz bus versions of the Pentium M, as well as by other processors such as the AMD Athlon and later AMD processor models with similar or more advanced versions of the same architecture.

http://pacoup.com/2009/05/27/pae-vs-64-bit-what-manufacturers-dont-want-you-to-know/comment-page-1

this option is compatible with any Intel Pentium Pro, Pentium II, III, 4, Core, Core 2, Core i7 and + processor, along with every recent AMD processors and Athlon series.

You should basically list these CPUs in the system requirements for your Linux distribution, or provide an alternate distro that has PAE disabled by default.

I'm looking at how much risk there is in setting PAE as the default mode in a linux distribution.

I'll answer from a different perspective -- how much end user benefit there is in PAE today. In my experience, PAE is a technology that even quite tech-savvy users don't know about or don't care about (or both)... For most sysadmins today, if they need more than ~3.5GB RAM, they'll immediately reach for a 64-bit OS.

So IMHO regardless of how little incompatibility risk PAE has, you should not enable it. PAE's use case has been completely taken over by 64bit addressing now.

Note however that NX is a feature that is marketed quite a lot, and that feature requires PAE (even Windows had to enable PAE with physical address space limited to 4GB in order to enable NX), and there is a lot of CPUs that support NX that do not support 64-bit.

Is there a benefit for systems with 32 bit cpu and 2 or fewer GB of ram? There might be. But I believe even a VIA C7M supports it, it might be possible that some very low end embeded CPUs might not, but they will not be running a standard distro anyways. I would say go for it if you have any reason to at all, unless you really target the low end, I doubt anyone will be hit by it. I think the early VIA CPU (the C3, perhaps the early C7?) may not have supported it (I know some were Pentium compatible, not PPro), and that is probably it.

I may be wrong but don't Linux distributions usually check for certain features during installation to decide which kernel to install? Ubuntu Precise, for example, has two kernels, pae and non-pae, and seems to choose the correct one.

Can't the kernel detect this during runtime yet? Ubuntu, again, for later kernels seem to have dropped -pae from their kernels. I don't know if this means they've dropped pae support (which I doubt) or it is detected during runtime.