I have made some more tests using the “spew” and “stress” packages, but it doesn't seem to cause the issue. So far, it mainly happened during package installation (thus causing the additional pain of leaving some packages in a broken state).
I have been unable to reproduce the issue with older kernels, but since I don't have a reliable way to trigger it, and that it took up to several hours to manifest, I cannot rule out this bug being present in older kernel verions, nor rule out it being a u-boot-related bug rather than a kernel one. (I am using the u-boot binary /usr/lib/u-boot/A20-OLinuXino-Lime2/u-boot-sunxi-with-spl.bin from u-boot-sunxi=2016.11+dfsg1-3 on both systems) Another annoying issue with this bug is that a full power cycle is needed for the µSD reader to be usable: even if I manage to trigger a reboot (opening /dev/watchdog for instance), the system won't actually be brought up again, and I have to physically switch it off and back. Not having any output set-up for this system, I can only presume that it fails to boot due to the µSD reader still being in a messed-up state and needing a full power-cycle to recover.
