I'm somewhere in between.  No high-power devices (that weren't self powered).  My power supply itself is fine.  My plug itself fell apart literally and I'm stuck using the power cable (I think of it as a tape-recorder AC cable which probably dates my childhood).

tar didn't choke the second time around.  I must have just hit the end of my DHCP lease right in the middle of my tar job by sheer dumb bad luck.  So you can use tar on an entire running file system.  Now I'm left pondering an implementation detail.  On recent kernels are the virtual filesystem directories generated by the kernel?  Or are they mounted into existing blank directories that need to be in the tree and probably have the right permissions to start.  If the latter that complicates restore, mostly to the extent of a logicless shell script.

Of course I think the ideal would be to have something where I could bootstrap the restore prucess in from Uboot somehow.

In the past few days, I've gotten by plug, got the Installer working, played around a little with the Ubuntu setup and installed a few basic packages.  Everything's running on the integrated NAND and I'd like to make a backup of my current setup to someplace else, ideally a USB thumbdrive before I start the really crazy stuff like getting a SANE backend up and running or software RAID 1 between thumbdrive and SDHC shared by Samba.

I've got a few ideas:
Tar-up the whole filesystem with a bunch of excludes.  I actually tried this already, but it didn't work for some reason.  Either I got it wrong or my network burped on me but it kicked me out of SSH and changed its IP address 16MB into the process.  That's still got me scratching my head and trying to remember how to spin off console-independent jobs without X.

dd the underlying partition: is that even safe?

Rootfs images.  That term keeps getting bandied about on the Wiki in ways I don't yet entirely grok, but it sounds like it might integrate nicely with Uboot which means I am VERY interested.

I'm sure someone has this figured out.

The DevKit Reference Design says the processor has two controllers, however, only one of them was actually laid out in the SheevaPlug board.  So yes such a thing could be built but it would take a new board.  Maybe a later product?