The fact that different survey subjects have different recall periods should not greatly affect the estimate of crude mortality rate or under 5 mortality ratio as long as the recall periods are not so widely different that different levels of recall bias are injected. Of course, another problem with widely different recall periods is that you are measuring mortality during different time periods in the same survey, so your survey estimate of mortality would be difficult to apply to the entire population. In addition, during data analysis you need to remember to calculate the correct number of time units in the different recall periods used in your survey sample.
In fact, theoretically, the recall period could be different for each group of survey subjects interviewed on the same day. Although the SMART recommendations recommend using the midpoint of data collection as the end of the recall period, it is easy to calculate a recall period for participating household by subtracting the date of the beginning of the recall period from the date of interview for that household. Then you merely sum the person-time units for each subject to get a total person-time for the entire survey sample.
Of course, if all of your survey subjects are from the same population and culture, the date at the beginning of the recall period should be the same for all survey subjects because that date should be a prominent holiday or other date well known in the population. The recall period should never be defined to the survey respondent as "90 days", "3 months", or some other arbitrary number of days. The SMART recommendations say that the beginning of the recall period "...could be a prominent date on the calendar, a religious festival day, a seasonal marker, or some other date of high relevance and common reference for the community." Because of the well-known psychological phenomenon called telescoping (see: https://en.wikipedia.org/wiki/Telescoping_effect), I strongly recommend that the beginning of the recall period always be such a prominent date and should never be some arbitrary number of days weeks or months in the past. For a slightly different discussion of measuring retrospective mortality in cross-sectional surveys, see page 42 ff. of the WFP publication "A Manual: Measuring and Interpreting Malnutrition and Mortality". It can be found at https://www.ennonline.net/attachments/924/45f6abc92.pdf.