I managed to solve it by running a separate goroutine (thread) which monitors the WAL size on disk every second: if it goes above the target size of 8 MB, my framework initiates the "slow down" mode where all reads and writes are slowed down by artificially calling sleep(), starting with 16ms and gradually increasing the sleep time according to a few heuristics. This allows the application to have small time gaps where no reads or writes happen and the checkpointer can actually proceed (the goroutine activates it manually). The slow down mode is deactivated when the WAL size is within the target size again.
I think SQLite in WAL mode is not really fit for production without this kind of hack.