When it comes to overflow-to-storage, I'd agree to push that responsibility to RegionFactory SPIs.
But we'd also need to consider the cost of storing & reloading such things vs. simply rebuild them: the 2nd level cache implementations would need some guidance or some weights to figure if it's better to evict things vs hitting disks.

Do you have precise data of what kind of things are making it "prohibitive"? We could also look at slimming the size of certain data structures, if we know which ones need special attention.