Thanks again, Kleinstein for the insight.
It seems that two of the three main temco contributions of the LTZ1000 reference circuit you mentioned comes from the internal heater related issues. Namely, the thermal gradient across the circuit and the temporal change of this thermal gradient.
In addition, the aging of the voltage reference, hence the annual drift in the accuracy, could also come from the prolonged heating of the circuit.
I could be wrong, but I think the humidity and the pressure effects are more pronounced for the resistive elements and to a lesser extent for the solid-state based zener reference.
Therefore, it may be possible that without using the heater function (i.e. turning off the internal oven) and opt for the external temperature stabilisation across the whole board (e.g. stabilize at slightly below 23 oC) may achieve a better overall stability?
For example, a typical room temperature change in a climatized room is about 1 oC. Combining this with a state-of-the-art LTZ1000 reference circuit tempco of 0.05 ppm/oC gives 0.05 ppm stability. However, if there is a tiny oven that can stabilize the temperature to 0.05 oC, then a 1 ppm/oC voltage reference can do the job just as well.
Does anyone know the tempco of LTZ1000 without the heater function?
Last but not least, regarding the long-term drift of resistive elements due to pressure, humidity...etc, it feels like the only remedy is to calibrate them as frequent as needed against primary standards.