The LM399 is overkill for a typical 4,5 digit meter. It might be useful as an external reference source to check the meter and possibly others from time to time, like once a month or to check how bad the ref. inside the ADC chip (usually not inside a µC) is. For a 4,5 digit meter a simpler ref. might be good enough. The cheap ref chips in plastic case may have a problem with the influence of humidity, but this a rather slow effect over many hours and days. The same may also apply to many resistors.
Bringing down the power of the LM399 by insulation has it's limits. The LM399 even without the heater active needs some power of something like 10-15 mW. If thermal resistance gets to high, temperature stabilization will not work at high environmental temperature, or the temperature needs to be set a too high a value. So it will get hard to bring the typical power to less than about 3 times the minimum power. If the long time reference does not need to run all time, power dissipation is not that critical any more. At low power and if the requirements are not that high there are also good band-gap references - preferable in a metal case.
Good sigma delta converters can give something like 2-20 ppm INL. Some may even allow for correction in software, as the shape seems to be rather predictable and fixed. This is well good enough für 4,5 digits, 5 digits and may even work for 6 digits. At the low ppm levels linearity is not that simple - even resistors may not be that linear any more. At least some of the SMUs already use sigma delta chips instead of a multi slope converter at 6 digits.
The multislope mechanism should not need that much compensation for dielectric absorption: only a small fraction of the resolution is determined from the change in voltage of the capacitor. As the capacitor is much smaller (e.g. 1 nF instead of around 500 nF in a dual slope) it's easily possible to use capacitors with low DA (e.g. NP0, mica or teflon). Still linearity is influenced by similar effects as in the sigma delta chips: voltage dependent charge injection, leakage, voltage dependent resistance of FET switches. This is not a surprise, as there is not that much difference between high resolution sigma delta converters and multi-slope converters. The difference to the classic dual slope is larger.
If you want the option of running on battery (even if rechargeable) I would definitely not use a multi chip, multi slope converter - this part alone would likely use way more power than the LM399. The sigma delta converters are really low power in contrast (e.g. 1 mW for an LTC2400).