seems odd for a buffer under run, as the play is not doing a great deal....
are you using 1.0.10?
one thing to be careful off, though probably not the issue is execution order , most of your midi and table/play objects are being executed before the mixer, I don't think this will have any effect, other than use the previous buffer, but always worth 'avoiding' these side effects.
also are the raw files all the full length of the buffer...
i think id try to narrow it down a bit, so its clearer what going on, in particular, perhaps you could sequence the notes, and see if this create a similar issue. It would be useful to remove the UI from the equation too... as this is sending data to axoloti which it needs to process.
(be careful with multiple notes, that you dont get distortion, so also consider turning patch saturation off)
also you could use patch/cyclecounter, to see how many cycles are being used... and if there is a large variance.
use max/min etc to track whats happening... depending where you place the cycle counts, you can measure different things,
sorry, hard to say anything really, as Id need the patch + samples... (its much quicker than having to generate my own, for a 'quick test', to play with ideas)