The problem was that our local coordinate system became very small

Contrary to what one might expect, it is the smallest representable
*positive* number.

Using -max() works because by IEEE 754, floats are symmetric!

With a zero nonlinearity