There's some truth to the argument that 1 mK resolution in air isn't meaningful, as there are many complicating factors besides the ADC. For example, with a relative air movement of 2 m/s, the dynamic pressure increase created when the air is slowed by hitting the sensor causes a rise in temperature of 2 mK. At 5 m/s, it's 12 mK. (Example calculated at 0 C.) I also expect that IR radiation complicates measurements. Etc.
However, radiosondes (and many other sensors) have 10 mK resolution. The self-heating should be below this figure and preferrably below 5 mK. According to my initial post, it's not obvious how to do even this with small sensors. Also, it's my hope that it's possible to measure temperature fluctuations in a chamber of a few liters with better resolution than 10 mK, even if not down to 1 mK. So I still think the question I posed is reasonable.
Yes, solids have a more well-defined temperature where 1 mK is a more reasonable resolution. For that case, the thermal coupling will also be much better so self-heating will decrease, giving margin for a larger excitation current and longer integration time.
Thanks for all tips and papers, I will have a look. I hoped there was already a schematics out there... oh well.