Hi Sander,
Personally I'd go for the i7.
Now the i7 runs circles around other processors when it comes to memory throughput since we work with large amounts of data this seems important but is it? Could it be that many processes work on chunks of data that fit in cache so that memory bandwidth isn't the bottleneck?
Of course throughput counts, but much more important is to have a *lot* of RAM. I'd consider 8 GB as a minimum. I recommend 16 GB. With a huge RAM everything runs faster (for example, you can accumulate many processing steps on previews) and you can do much more complex things (f.e. large wavelet transforms, especially HDRWT).
Are there plans to use GPU acceleration? I've read a bit about CUDA but I'm not sure it's available cross platform.
My advice is: buy an nVidia graphics card

Not in the short term, but PixInsight will use GPU acceleration via CUDA. Hopefully this year we'll have some core routines well implemented through CUDA. Convolutions and FFTs are the main initial candidates. Then depending on how things evolve we'll implement more sophisticated algorithms.
CUDA is cross platform. There are excellent implementations for Linux/UNIX, OS X and Windows. I have already studied some CUDA example code, and Oriol is also working on this.
My main problem now is that I have to update my main Linux workstation to install a CUDA-enabled video driver, which implies a recompilation of my Linux kernel, along with several key packages (VMware), monster backup copies, etc. This is stopping me because right now I need a fully operational machine all the time, and these changes would leave me out-of-business for several days. As soon as I can afford a full software update, I'll begin working on a port of some PCL routines with CUDA. I'll keep you informed.
The machine will most likely run 64 bit Vista.
It won't deserve that

Seriously, you'll really hate Vista. I'd consider waiting for Win7, and use XP in the meanwhile. Or, much better, why not install a good Linux distribution instead?

I'm not sure yet about the disk config. I realize that PI likes to have multiple drives for swap files. I don't think I'll accommodate that right away but use fast SATA drives and RAID for general speed and data protection of all applications.
I use RAID 1 for the sake of security on all Linux workstations. RAID 5 and RAID 10 are good choices of course, but they need more drives, and nothing is more secure than a simple RAID 1.
Use a good software RAID implementation. This is also a good reason to use Linux

Beware of mainboard RAID utilities; they are actually software RAID implementations that use comparatively slow processors, even the best ones, and if your motherboard crashes, you may have serious trouble to restore your drives due to proprietary RAID software. If you cannot afford a server-level dedicated RAID card, use software raid. Your i7 processor can perform all necessary RAID operations much faster than a mainboard based RAID. For example, I have two 750 GB SATA drives configured as RAID 1 with mdadm on Linux. I've never seen a processor load above 1%, even during intensive disk I/O operations. With this configuration, I have a consistent write speed of about 475 MB per second, and around 1100 MB/s for reading.
Buy two or three cheap, relatively small SATA drives with a good cache size, and use them only as swap storage drives. PI will perform undo/redo disk operations almost as if they'd happen in RAM

Spreading out the 'investment' keeps prying eyes (wife) at ease you know
Oh yes, of course. And you can invest more with a fraction of the total "pain"
Good luck with your new machine.