2. What exactly qualifies as a "Linux server"?
-Obviously, you are including only systems where it is fully supported by the hw vendor and available with Linux installed; a reasonable move, though fully supported by the distribution would be a better criteria if you're only interested in the capabilities
-Does it have to be Linux as the default configuration, or only option?
Because HP has HP/UX and OpenVMS on anything Itanium, and IBM has AIX on anything POWER (except for a few budget configurations that require an open-source OS like Linux or the OpenSolaris POWER port), and SGI and Cray offer Windows HPC as well as Linux, and Oracle has Solaris, OEL, and "Oracle VM server" (Linux/Xen). I am not aware of any Linux-only vendors. But then, the only current UNIX-only vendor I know of is Apple.
I'm trying to figure out what makes you exclude the Superdome, because I don't see the logic there. It's offered with RHEL (the industry standard), and using someone else's OS with your hardware is standard in the Linux world.
3. Ted Tso is very much a desktop developer, and there are thousands of kernel contributors (I mean that literally). It is not reasonable to suppose that he knows about the configurations that SGI, Cray, and IBM kernel developers have access to.
Each vendor has its own favorite approach to scaling. Oracle scales first by adding more cores per box. IBM scales primarily by boosting the clock and adding blades. SGI scales by adding racks/nodes, and likewise for Cray; they found in the 90's that the most space-efficient route was to put fewer sockets in a smaller system, then put more systems in.
So you're asking why nobody uses Oracle-style "scale up" configurations with Linux. My guess is that everyone else figured it was a dead end: if I have an M5, I can go up to 32 sockets and that's it, while if I get a comparable multi-node system, I can add more nodes and get much farther.
Why do you think that the kernel (NOW, not as it was in March 2006 when kernel 2.6.16 from SLES 10 was released) can't handle 8+ sockets when it has been used for 100-core chips like Tilera makes, and the XLPII (MIPS64, 4 threads/core, 20 cores, and somehow they can fit up to 8 cores per socket) is primarily used with Linux?