Hi Lars Corneliussen, basically what happened here was that the fix you'd applied got lost when I refactored. I noticed that, and reinstated it, and should now be working as before.
However, I noted that this might not be the optimal solution in the long run - the scope is based on where it is instantiated rather than where it is used. Also, they are named as "caches", but they seem to be behaving as a way to collect information during a resolution run, that can be accessed after the resolution run, but are not to be used by subsequent resolution runs (hence the problem when retained where they accumulated information from previous resolutions).
I don't think there is anything actionable here now - just wanted to add some information in case it's helpful for future reference.
This is completely separate to
NPANDAY-599, where we've also been discussing caching (which actually are keyed resolution caches for other scenarios), which I confusingly referenced this change before I understood what was happening there.