Hi,
the Reference Column Splitter always keeps the columns from the reference table (output 0 - top) but, logically wise, I’d expect the reference columns to be split “away” (output 1 - bottom).
That unclear behavior - who reads the fineprint 
- results in messy workflows (crossing connections) and has the potential to cause data issues.
In order to improve understanding of the nodes functionality I want to suggest adding an option to steer the referenced columns either to port 0 or 1.
Best
Mike