Unfortunately CTOL and TOL seems not to be relative and so size matters 😉 We had that discussion a short time ago with a similar problem with not suitable results using genfit() with units in Prime.
To compensate for the nano in your example and given that the dimension of the function you wish to find the roots for is volume you would have to set CTOL to 10^-30 to get the same results as with CTOL=10^-3 without units (or with units meter). But you can't set the effective value of CTOL to anything lower than 10^-16.
So there is no other solution than to give the advice to avoid units (other than base units) in conjunction with approximation algorithms which rely on convergence even if you are using Prime.
Its interesting, though, that Find() is not affected. Unfortunately in Prime we cannot change the algorithm used but we can do so in MC15. And here are the results:
Levenberg-Marquardt: 10 nm, perfectly OK
Conjugate Gradient: 1 nm; simply returns the guess regardless of CTOL
Quasi-Newton: 1 nm; same behaviour as Conj.Grad.
(Linear: 0 nm)
