Back in May, I tried creating unit tests for the cache's new stateful features.
This scaled poorly, which later led to the creation of the Barry/RPT framework.
I'm throwing away those failed unit tests, but I'm keeping the few bugfixes I
managed to whip up while I was coding them.
The only special mention is "separate cache_setup() into two steps."
This is needed because all involved processes need to work in the cache
directory. Ensuring the directory exists and has been chdir'd into before
subprocesses are spawned is a simple means to achieve that.