Part of the explanation may be that when you run an idealized case, you do not have input for the nest; hence the model interpolates everything from the coarse domain to the nest, including latitude and longitude. If you set feedback = 1, it will cause nest data to overwrite the coarse domain in the area of the nest, including XLAT and XLONG. This is necessary since the calculated values from the nest are replacing coarse domain values (because of radiation, for example). Unless the interpolation is really bad, the odd ratio should do reasonably well, as there is an overlapping grid on the nest and coarse domains. For an even ratio, since there is no co-locating grid, there will be some averaging done when feedback is on.