Hi,
After investigating carefully, I have not spotted any particular problem with the FFT model. There could be some numerical issues who can cause this effect.
I have noticed that for example the fit fails to get the right value when injecting a fake signal.
I would try to re-parametrize the parameter of interest “ratio”, to be in a more sensible range (e.g. [0,10] ) by rescaling it by a 10^-5 or 6 factor.
Lorenzo