My previous attempt at averaging out the charging noise involved taking 10 ADC readings, but at the weekend I learnt that this is unlikely to work -- 10 ADC readings is very quick and my noise is at 50Hz. I'm now averaging for 160ms or thereabouts. Averaging 8 cycles of the noise is probably much more than I need, but it solves my problem nicely. It is causing problems with the scan time.
The firmware is continuously monitoring the cell voltage, and when a request comes in it will have to wait for any ADC averaging that is already going on before waiting another 160ms for a new reading to respond to the request. Since we take two ADC readings to measure the voltage, and we also ask for the shunt current, it takes about 10 seconds to scan 10 cells (I now have two BMS boards up on the bus). I will improve this in two ways, the first is to cache the most recnet values so that they do not need to be regenerated when a request to send them is received. The second will be to interrupt the currently executing ADC reading to return a cached value. It may be possible to do this latter change purely in the receive interrupt handler, if it isn't it will likely involve a significant design change to the firmware's main loop.
The above graph shows quantisation of about 9mV. This is unexpected with so many values averaged. I will confirm the average calculations by streaming the raw data back across the bus. I'm currently running at 9600bps, I may increase to see more of the data. Charging current was between 5 and 10A.