Hardware Update March 2008:
Got a G4/1ghz sonnet processor card. The G4 cards present many more difficulties in the process of getting them to work properly than did any of the G3 cards I owned.
First, Linux offers no way to enable the L3 cache on the G4 card. It assumes that it's been enabled in open firmware and since L2CR vanished from /proc in 2.6.16.something you can't get at that directly any more either. To overcome this, I added a kernel parameter to set the l3cr on boot. A patch is here. It works on 2.6.20 or so and above. My boot command line now includes this l2cr=0x8000000 l3cr=0x9f020300. Unless you are using this exact card your parameters will probably be different.
Second, I had a devil of a time with stability on the 9600. The onboard L2 has been disabled for a while so that was not the problem. The system would simply freeze for no reason. It also ONLY worked with the onboard mesh scsi or a Adaptec 2940 card. The 66 mhz IDE card I got from OWC caused corrupted data when used with the G4. I tried dropping the base (bus) freq of the G4 card by replacing it's xtal can (50 mhz) with a 42 mhz one but that was no help. As a side note, this often helps with G3 cards on the old-world Macs, dropping the bus speed to 46 mhz or so and pumping up the multiplier can result in a much more stable machine.
Sonnet advised that the G4 cards were much more picky about your PCI cards. Since I have so many PCI cards installed, I want through an exhaustive process of replacing them with other cards to try and find some stability. I never did get the 9600 working properly (except for when it had only a video card and a scsi card in it). Reluctantly I went back the S900 chassis (I was reluctant since the S900 ties all 4 bottom slot IRQS together whereas the 9600 has separate IRQS for all of the slots).
Third, the S900 was picky too! I think I might have just achieved stability in my current setup. The G4 card back up to 48mhz/960mhz (probably would be OK at 50/1000 but I'm not taking any chances). Also, I tried hard to avoid ganged interrupts on the bottom 4 slots. The machine didn't lock up but occasionally it would hang for a while (presumably stuck in some IRQ loop) and then recover. This was with 2 tulip cards, one for internal ethernet and one for external on the same IRQ in the bottom 4 slots.
Fourth, Open Firmware in old world macs doesn't work well with more than one PCI card with firmware on them, ESPECIALLY with the G4 processor card. I couldn't use my Radeon 7000 with the G4 card no matter what I did, it wouldn't initialize.

What I've got running now: PCI cards top to bottom, ATTO scsi express PSC (flashed from apple firmware to ATTO firmware or it wouldn't work), USB2 card, open slot, intel pro100+ ethernet, Rocketport PCI, Rage 128. I'm using the internal MACE ethernet for connection to a DSL modem and the internal 8530 serial port for monitoring the APC UPS with APCUPSD. A Microsoft Natural Keyboard Elite and a Logitech Trackman Wheel to save me from carpal tunnel (works great).

Hardware Update November 2008:
My patch for L3CR made it to the mainline kernel with credit! Wheee, my 3rd kernel contribution!
Still using the S900. Finally figured out that the ganged interrupts in the lower 4 slots are not reliable. They crash the kernel when the interrupts are generated faster than, say a USB 1 card. The only cards I have in the bottom 4 slots now are ones that don't generate interrupts. (the rocketport card is polled, I'm not using interrupts on the Rage 128 card and the other 2 slots are open). This forced me to use a usb2->ethernet dongle to get the second ethernet port without using one of the bad slots. BTW, this seems related to the G4 somehow, the G3 was rock stable with the cards in the lower slot and it doesn't matter how low I drop the base clock speed of the G4 card, stuff that generates lots of interrupts in the lower slots simply crash the machine.
What I've got running now: PCI cards top to bottom, Promise Ultra 100 20268 IDE card) which has no Mac firmware and as such is NOT bootable), USB 2.0 5 port card (with asix 100mb usb 2.0 ethernet dongle amongst the other USB stuff), open slot, open slot, Rocketport PCI, Rage 128. An Adesso natural keyboard and logitech trackman wheel round out the input devices. I boot from a small scsi drive attached to the internal mesh scsi controller switching to root on a 185g ide drive on the PC controller then spin down the unused boot drive with scsi-spin.

Using an ASUS PCI 200 (ncr53c8xx) scsi controller resulted in a 50% speedup for the drives though it can't be used as a boot HD controller. The 2940 was even faster and has been used successfully by others as a boot controller.

I recently installed LinuxPPC 2000 Q4 edition (Halloween) on the 7200 and it even boots with yaboot even though that is not supposed to work on old world macs. Very interesting! Figured this out, it uses miboot! The 7200 is currently running the MN amprnet gateway at Augsburg college, it's running kernel 2.6.14.

Ins and Outs of the 7500:
The 7500 has exactly the same case as the 7200. It is about the same speed (stock 100 mhz compared to 90 mhz of my 7200). There are some major differences though. First, it has a real synchronous SCSI controller (the linux driver name is MESH). This works much better with PPC linux. I can even burn CD's at 4x without any hassles now. With a WD ultra scsi drive I get 3.5 megs/sec on writes and 4 megs/sec on reads with iozone. The tape drive is connected to the external scsi controller (which BTW also has an internal connector). The external controller is the same as the primary controller in the 7200, a NCR 53C94. The reason I connnected the tape drive (as well as the CDRW drive) to the external SCSI bus is because the tape drive was limiting the internal scsi bus to 5 mbits/sec. With the tape drive moved the hard drive now runs at 10 mbits/sec.

Unfortunately, the internal MACE ethernet STILL doesn't work right. I would like to add a nice ATI 3d rage pro agp 2x card to the machine to get accelerated X but can't until I free up a slot by getting the mace to work right. Another option would be to go with the experimental XFree 4.0x Xfree86_FBDev server which has been patched to provide some acceleration with the 7500's control video chip. Unforch, the controlfb video driver has been unwilling to sense the full 4 megs of ram in the 7500. There is a fix for the 2.4.x kernels and though I have other unresolved problems with them, I'm now running 2.4.x full time. My current kernel will run on the S900. It includes support for various amateur radio drivers as well as the Rocketport Serial, Tulip and ne2k-pci ethernet drivers.

Outstanding Issues with the 7500:
MACE is unreliable (maybe fixed now?) and the 53C94 external scsi doesn't support disconnect and hangs on the slightest error. (Really was never fixed).

Interesting things about the 9600:
It's got 6 slots and a second Bandit PCI controller! Otherwise, very similar to a 7500.

Interesting things about the S900: Like a 9600 but with a DEC PCI controller for the lower 3 slots instead of a second bandit, ties the interrupts for the last 4 slots together (boo hiss). Has a second processor slot for a custom 604 (not useful for me). Has less ram slots (8 as opposed to the 12 in the 9600). Had onboard ram AND cache which needed to be disabled to run with G3/G4 cards. It's open firmware is a bit funky, doesn't report that it's actually an apple. Nice small case (compared to the 9600), runs cool. Interestingly, neither the 9600 or the S900 will let me see more than 768 megs of ram in linux though more shows up in Mac OS9.

I'm using the XFS filesystem.