Some process activity instances get missed when there is a high amount of data generated while the import is running concurrently and has progressed to the tip of the current time.
In order to decrease the chance of missing concurrent writes in the engine the latest timestamp saved to continue in the next import cycle should get backed-off by a grace offset from the current time.

Local Load-Testing with 10k-50k generated process instances showed that 2s appears to be a reasonable offset. The cost of this mitigation is reprocessing already imported entities and can be considered acceptable.