Overall: very nice! The implementation seems clear and to the point, and (even better) the test suite is great.
I have left a number of, mostly fairly minor, comments on GitHub. I was a bit bothered by the large number of quite similar versions of countsToMagnitude/countsToMaggies. I see you've already commented that they are all in active use or have a use coming up, but still — I wondered (but have not checked) if it'd be possible to refactor the callers slightly to avoid pushing all the possible alternatives into this class.
Also worth noting is that I've no real experience with pybind11. I looked through your code and it seems fine, but it might be worth asking one of the experts (Russell, Krzysztof, Pim, ...) to glance through it in case I'm missing some subtleties.
Specific points you raise above:
testPhotoCalib.py fails the memory test, triggered by testPersistence (you can show this by unittest.skip()ing it. I couldn't figure out why. Ideas welcome.
I think my comment above should cover this.
I had to make several changes to BoundedField to make this work: those are in separate commits. I didn't file tickets about those changes, but maybe I should have, to capture the story points?
I have no strong feeling on this — I'd say they're fine as they are. Maybe Simon Krughoff cares?
Please check my math on the error calculations. I think I got it right (thanks Mathematica!), but extra eyes are welcome there.
I tried to work this out on the proverbial envelope back, and I got a log(10) in a different place. I may well be confused, but perhaps you could take another look and feel free to ping me if we disagree.