I would also like to compare the performance of a few approaches
(1 file per sys.path hash using INI, JSON, and sqlite; one file per
entry on sys.path using the same formats) using a significant number
of plugins (~100?) before we decide.
If you can manage it, you'll want to run at least some of those tests with the plugins and their metadata mounted via a network drive. When the import system switched from multiple stat calls to cached os.listdir() lookups, SSD and spinning disk imports received a minor speedup, but NFS imports improved *dramatically* (folks reported order of magnitude improvements, along the lines of startup times dropping from 2-3 seconds to 200-300 ms).
I'd expect to see a similar pattern here - inefficient file access patterns can be tolerable with an SSD, and even spinning disks, but the higher latency involved in accessing network drives will make you pay for it.