Summary statistics for short [1 run @ 64 windows] (1) and longer [100x] (2) show rather similar responses, where channels 6&7 are clearly noisier, while all others are in decent agreement.

It becomes clearer in the larger statistics that the distribution looks bimodal.  This could be due to amplifier chain instability/oscillation.

As a check of trying a common mode subtraction, did ch1-ch0 (3) and ch2-ch0 (4).  In the first case this perhaps makes sense, if the signal is 500MHz, the phase difference leads to the difference seen.  Should try to plot this modulo period and see if forms a periodogram.  For channels that should be in phase, the improvement isn't as much as would think, so this needs further investigation.