Still, this approach was for a variable number of traces, which requires reading all traces in TTree, thus I guess out of the question unless there is some way to implement this ~2D variability in a more optimal way.
Well, I wanted to try the all-detectors-in-the-experiment approach from the stard and I’ve started coding it. While the size impact shouldn’t be big indeed, I wonder if having, for example, 200,000 traces per event, most of them empty, will have impact on performance.