Do note that this is 15,786 sequential appends, though. FDB responds success to a client when the commit is sync’d to the transaction log, not when it’s applied on a storage server. There’s up 15,786 random writes that still need to happen to disk after the 0.63 seconds that you’ve recorded before the written data can be read. These writes will actually be repeated, because they’ll be recorded both in the sqlite WAL and the sqlite btree.
I believe the topic of embedded FoundationDB has come up before, and was cautioned as unlikely to be a fruitful direction.