A rough guess without knowing any better, would be that their SQLite configuration isn't flushing to disk after every write or similar.
Maybe they ran their entire benchmark in a single transaction or something?
The "on-disk" test was really in-memory too, thanks to various layers of caching etc.
Sounds like they were testing it with a small enough amount of data that the disk version could just work from cache and didn't actually have to go to the disk