OK, an update. At this point I'm about ready to cut a branch, commit, and submit a pull request. The driver works, though I still have some issues with it.
I'm now using an appropriate Zener with a normally-distributed noise voltage source added as the basis for the noise generator: the ad-hoc random current generator is gone.
I also abandoned the low-gain hack of the op-amp after I discovered that getting the oscillators to run at the right frequency and voltage range with that op-amp hack was no easier than with the standard op-amp model with normal gain: I was only getting increased numerical stability because the lower gain kept the oscillator running in a narrower voltage range. Instead I went back to the method of putting a frontier on the Schmitt trigger op-amp's output, which seems to remove the numerical stability problem. Some slight tweaking of the Schmitt trigger op-amp's resistors corrects any change in frequency and voltage range. I no longer adjust the oscillator capacitors.
I removed one of the op-amps in the final amplification stage, since it was effectively unused. Its only purpose in the real circuit is to invert the other op-amp's output so that the final power amps can get opposing signals in order to drive the speaker "push-pull" style.
I took out the reduced accuracy and tolerance parameters for the solver. They didn't improve speed much and reduced sound quality.
Unthrottled performance with static solvers is still around 170%-220% of full speed on my machine, depending on sound workload.
The sound seems higher in treble and lower in bass than what I hear in the few videos of real 280-ZZZAP machines I can find, and the "skid" sound is less distinct than on those machines. Oh well. That may be a difference in the machine's audio "front end" (is that the correct term?). There also may be some differences in component values between the schematic I used and real PCBs. It would be nice to compare with real PCB photos, but I can't find any.
This driver should also be able to support Laguna Racer and Super Speed Race, but there are some component value differences (and circuit differences for Super Speed Race). I don't know how best to handle that: duplicating the whole netlist with components changed seems clumsy. Is there a way to parameterize or have conditional code in a netlist?