Why would you need/want QEMU's temporary snapshots? Just create your base image, and each time, use a copy (or cow copy) of that - can be on/in temporary storage, or even in RAM, just apply any relevant fixups to it before firing it off (shouldn't be too identical, e.g. host private keys, UUIDs, etc. - those should be unique each time).

Ideally the backing VM is on disk and I'm running the immutable VM on tmpfs so I can start a new VM frequently without wearing out my SSD

Should be pretty easy to do. Have your base image, do cow or unionfs or the like atop that, and that upper layer can be in RAM, needn't be on drive, do your fixups there, and launch and run. Only bits that would be changing are the bits to initialize and run the fixup, and any (non-persistent) changes within the VM after that. And if you're going to repeatedly relaunch same VM from same initial state, then there's even less that gets written, and most of it would be in RAM.

I use wrapper script/programs for some of the more complex libvirt operations I semi-commonly do, e..g creating a fresh VM, or doing live migrations of some VMs. Wouldn't be too hard to write something like that to do VM(s) as you describe.

And I don't see any particular need for "snapshots". Base image would be clear cold shutdown image after initial built VM is shut down - no need to grab snapshot of a running VM. Just suitably clean up and reinitialize it (e.g empty out log files and other stuff not of interest), then you have your main base image, and can go from there to, e.g. reinitialize for VMs you fire off from that (e.g. update host private keys and UUIDs, etc., and then launched based off of that). And that reinitialization can be done in cow or unionfs layer before launch, and that storage thereof can be in RAM, with only the backing store on, e.g. SSD.