This option "concatenate input files before process" put all the input recordings one after the other, and passes it to the coherence computation function as if it were one single piece of continuous data. The separation between the epochs are ignored, leading the cross-spectrum ("PSD") to be estimated from segments of recordings that may include jumps between epochs. Definitely not ideal, but could be a solution when trying to estimate coherence at lower frequencies with short epochs.
See this tutorial for explanations: https://neuroimage.usc.edu/brainstorm/Tutorials/ArtifactsFilter#Evaluation_of_the_noise_level
Ideally, you would use only continuous block of recordings that are long enough to estimate the coherence correctly without having to use this concatenation "trick" (steady-state, resting-state, or at least brain processes that are long enough to have some sustained oscillations). In that case, you would use either the option "save individual results" or "saved average conductivity matrix".
The screen captures and description of the options in the connectivity tutorial do not match your recent updates... Could please update them?
The way the time is handled should be explained as well (who are the signals chopped, what's the duration of the estimator window, etc)