We don't have good tests for end-to-end cache performance. It would be good to add tests that check that cache files written in a normal run with --cache-fine-grained are then successfully loaded in a subsequent daemon cold start. The emphasis here is on end-to-end, we have tests that use cache but they don't exactly emulate what happens "in real world". An additional argument for this is that code in build.py responsible for cache loading is quite complicated, so it is better to be sure that we don't break it.