In the most interesting voltage range (3.1 to 3.4V for LFP) I mitigated this by using differential measurement. The divided voltage is compared to a 2.5V reference and the difference is multiplied with the internal 20x gain stage (attiny 26 offers 8x, 20x, 32x). For NMC with its steep voltage to SoC curve that is not even necessary though.
News
Today I played with the BMS a bit more after it has been resting all winter. The comm connectors were slightly corroded and I had to wiggle them quite a bit to get communication going again. I also found the voltage accuracy has suffered, probably also due to corrosion. So I will now lower the divider resistors by 5x. That adds about 40uA quiescent current on the bottom cell (30, 20, 10 on the others) which I think is still acceptable.
I also found that JLCPCB now offers to populate connectors. I will do a small run there and replace the boards on my test system with it. Can't make any more head units right now because all variants of the STM32F105 are now sold out
