The R + 9R + 90R + ... implementation is really a much used solution. Some companies sells the resistors in a complete kit.

And as Erik noted, the temperature coefficient has to be taken into account. If you measure for several hours, then a lot can happen. Traditional high-quality equipment normally uses owens. You have to wait from minutes to an hour or more until the critical components has reached the expected temperature.

And even for equipment without owens, you normally will see a change in measurements during the first seconds to minutes because of changes in temperature and all capacitors reaching a stable state etc.

Some Agilent bench multimeters gives their specifications after a 30min warm-up, and when used within 18°C to 28°C which seems to indicate that they either have owens or software to handle this 10 degreee ambient range. I would expect that you can somewhere find a similar warm-up requirement for Fluke high-end multimeters.

And when you do read the documentation for instruments, they may claim x% + y digits at full scale or they may alternatively specify the error as % of full range + % of actual measure. If they specify 0.1% + 4 digits and full range is 4000 then they can have a measurement error of 0.001*4000 + 4 = +/- 8 ticks, which would be +/- 0.2%. After switching to the next higher range, the measurement could then (for some instruments) be 400 +/- 8 ticks which is an error of +/- 2%. This is not too uncommon figures when measuring DC current or DC voltage. Switching to AC or resistance normally drops the precision a lot.

But the above means that you can't just settle for 0.1% at full range - you need way better than 0.1% at full range, to get 0.1% at 10% of full range, i.e. directly before/after a range change. Time to take a closer look at them 0.01% resistors :)

Both Agilent and Fluke can supply you with a sub $1000 bench multimeter managing your precision, but not your speed. I would expect them to gain their precision from using dual-slope conversions, to allow them to cancel errors with the voltage reference on every single sample. That means that the quality of the shunt resistors will be the main error factor. In your situations, you need to know the value of your shunts with an error way less than 0.1% since you need to make room for the ADC errors and for fractional-scale measurements. It makes a big difference to try to auto-cablibrate once before a multi-hour run, and performing a dual-slope measurement for each and every measure. With the dual-slope, every single measure will cancel errors from the voltage reference. The two slopes are just ms apart, so there will be no significant temperature change or supply-voltage change between the measurement slope and the reference slope.