Ok, so it seems that this 2048 characters limit applies only to text given via UI. When passing parameters via the notebook activity in ADF, you're essentially embedding the value in a JSON payload sent to Databricks. So maybe there's a some kind of limitation regarding accepted payload size on the side of databricks? 

So you meet payload requirements from the side of ADF, but hitting limitations on databricks side.

As a workaround, can't you split output into smaller pieces? Or maybe write an output to a parquet file that can be later consumer consumed by your notebook.