This is very interesting, power cycle suggests boot device read error which is spurious. Due to a transient in the power supply as you are positing. This is a very large deployment. What is the power into the board ? Have you managed to recreate this in the lab ? As the saying goes - at scale things fail.
Without bench recreate this it would be hard to verify, are theres sources of noise around the board that can affect the SPI bus data integrity during boot ? I had this issue when using the Onion in energy monitoring devices and connecting the monitor to the Onion over an SPI bus and sharing the power supply. Occassionally when switching inductive loads, particularly LED or Fluro lamps the Onion would crash with this behaviour.