Bacause poly resistors' conduction depends on both
simple doping, and grain boundary effects. If the
latter dominate tempco, then you could see negative.
The foundries do not, cannot control tempco based on
room temp WAT feedback and often make no attempt
at all to do so; likewise modeling can't do much for you
when the direction, let alone magnitude of drift is not
known beyond a few initial lots' worth of data.
But poly will give you density and low capacitances,
while well and S/D resistors are less dense and have
significant parasitics.