In such cases, it may help to run the FORS_calib recipe with modified parameters. Changing the parameters cdegree and cmode from their default values (especially setting cmode = 2) should help in allowing for a better detection of the edges of not well separated spectra. Varying the cdegree and cmode parameters and comparing the extracted spectra with the ones expected for a "good" (i.e. non-overlapping) slit until a good match is found may be the best way to extract such spectra. However, ultimately, it may not always be possible to extract overlapping MXU spectra in an automated fashion (which is why this is strongly discouraged in MXU mask design).