Thanks for the replies. I may well have to go the chopper route, though on previous looks through the parametric tables I've ended up excluding them because of other factors, however this has been limited by my understanding of various things so it's worth another look. Building a little oven might well be the best solution since I think I'm going to need one for my voltage reference anyway. With an overall amp input drift of 0.05uV/C I would get 10C room within my ideal error budget of 500nV, 20C max.
Take the LMP2014 - I got excited when I saw that until I saw "differential input resistance: 9M" I later realised this isn't so bad since the input is entirely common mode, but what is the common mode input resistance? The datasheet is strangely quiet about that. Also input offset current is 6pA compared to bias current of -3pa. As far as I understand, offset current is the difference between bias currents, so this just doesn't look right.
Re LMP2021, Ib is too high! I need better than ~10pA. I've looked at a few of TI's chopper and auto-zero amps, they've been a case of nearly, but not quite. I need to spend more time looking at Linear's amps though I think, however a lot of those have Ib compensation so the current although tiny is +/-, which is no good for me unless it's really tiny. Same goes for Intersil and Microchip amps, both offer amps ideal in every other respect. Bit of a bugger, really. (Eg, ISL28107, Ib is 15pA typical, but the min value is -300 and the max value is +300. Too much chance involved)
The diode idea is where I had started going but I hit a kind of mental brick wall there. If I can find one with a suitable TC I wonder if it would be adequate to just put it in series or parallel with the offset null pot? There's already a constant current planned for that so it's part-way there. Doubt if good enough by itself but might complement the oven idea so I can use an amp with more average sort of temp drift?
Have to get ready for work now. Sunday afternoon too. Boooooooooooo.....