While working on some tests, I was having a tricky problem in a test suite. Eventually I tracked it down to an interaction between tests. I suspected the test library, but once I tried to make an isolated test case, I realized the test library was working fine. It turns out it was the server’s request cache. The fix is to clear the cache between tests.
Not needed for this PR, though I’m adding it to this branch because it conflicts with this change.
The old LRU implementation stored a list's indices to reference items in that
list, but deletions from the list made indices point to the wrong slot.
Functionally, it meant that deleted slots were not guaranteed to be the oldest
slot.
Using a linked list fixes that.
