Compared to the hidden overhead of the actual output that trace enables:
1 small layer of encapsulation will be difficult to measure.
Recent experience with the 4GL Interpreter gives an Insight:
I had a small loop that would increment a counter and then show it, quite similar to trace( ).
I thought the singing and dancing through the Data, Operator and Noun stacks had significant overhead. Time spent doing the output itself was about 50x the Time running the Interpreter. I approached the updating now by only showing about 5 times a second instead of every time. After all the numbers would still be a blur to us Humans.
So I don’t see a Performance problem.
As far as Purity (perceived Technical Debt), I suggest a Pragmatic approach here and just go with it and worry about the likely other much larger Dragons you will need to slay later 