It seems the current implementation waits for the duration of the polling interval (100ms by default) between every single batch of events.

Proof: when enabling the TRACE level of logs, I see things like this right in the middle of processing, when the outbox is definitely not empty:

It's a mistake, and this can dramatically slow down event processing.

When there are more events to process, the next batch should be processed immediately, not after having waited. This was the original intent.