At which point, keying all accessing to

compile_classpath

by subgraph instead of by target actually doesn't thrash the cache much