I noticed that a feedback-loop would recook from scratch when I increase the iterations amount, instead of smartly building up from the last iteration cache. This a big waste of time.
Do you know why such behavior?

I've managed to fix this with some python and geometry data parameter used to store the last iteration cache.

Then I realized that I could use a simply Sop Solver as template for a feedback loop. Sop Solver is handy bc comes bundled with caching functionality already.
The only problem with a sop solver, is that I haven't found a way to control its generated cache with python. It seems impossible?
For example, I want to clear all the cache before frame 10.

It recooks because the result may be different, for example if each iteration also uses numiterations metadata

oh I didn't think about that eventuality, thanks