This is not really a bug, but under non-ideal conditions (large Hamiltonians, non-sufficient accuracy settings), the numerically calculated eigenvalues wrsk in the COMPUTE routine can have a small but significant imaginary part. This can translate to peak frequencies that have an imaginary part if the DigitalFrequencyResolution->True setting fails to "snap" the peak frequency to the frequency grid. In that case the FT might fail.

I would recommend taking Re[wrsk] of the eigenvalues and/or print a warning if the imaginary part is significant.

Please see attached your RotationalResonance.nb Example, that I forced into the above behavior.