Clear the set of known-traversed objects before building the graph for the next commit's root-tree. Otherwise our tree graphs will be incomplete, as objects referred to via multiple commits would only be stored in the graph of one particular root-tree.

If self.counts is not initialized then failing tests may produce an additional error about self.counts not existing, rather than failing on actual test assertions.

This should reduce memory requirements significantly. Storing all trees of all commits with all tree entries in a single igraph uses too much memory to load repositories such as git.git.

The aim is to reduce the memory footprint of object graph construction