You can use SQL as a document store, you don't have to use joins.

MongoDB queues inserts into memory before flushing it to disk. This means it's not durable in case of a power outage before the data is flushed. Turn on the safety of flushing-first and it's more or less doing the same stuff your sql software is doing.

Is is, in fact, much much slower than both SQLite and Postgres. It's even slower than a fully indexed hstore in postgres! (which is pretty much identical to mongo's data model, except consistent, durable, with transactions, etc.)

That doesn't make sense. Relational databases use relational algebra (which includes joins) to model data. The mere fact that you model it that way doesn't necessarily have to affect the physical representation. True, there is no such thing as an actual RDBMS on the market, but that's hardly Dr. Codd's fault. :-)