I wound up spending most of the holiday weekend lashed to a computer but not on the Interwebz.
While taking Real World performance data on a bit of hardware last week, my colleagues and I discovered some odd behavior that was not completely unexpected but had not been predicted by the original modeling of the circuit. I’ve spent a large part of the past few days refining that model. I’ve improved it to the point where it will do the same kind of tricks as the Real World circuit, but not yet in exactly the same way.
It turns out that a couple of the components are more temperature sensitive than assumed in the original model. The warming isn’t global. It’s just a local phenomenon, and the unexpected performance has turned out to be a feature rather than a bug.