Put a cache on split()

I found the split function was taking up a fair share of CPU in some workloads. I got an order of magnitude speed growth by adding a cache. Might be able to do something as simple as an object map of names to arrays that gets cleared every 15 seconds

The issue names split() but no file or test; first locate that function and inspect how it is used in the tree structure. Measure the current workload and decide how a name-to-array cache with periodic clearing should behave, then verify that repeated splits reduce CPU without returning stale results.

Written by the indexing model from the issue text.