readtable and friends do their best to figure out what the data types are by pre-reading and analyzing the initial content of the file -- what causes the difference without using additional input parameters or an explicit import object is where in the file are the missing data; if they are relatively early, then the algorithm for data types will be influenced and may, as you've discovered, return what are thought to be numeric columns as strings.
Unfortunately, the exact conditions used to determine which is which are not documented, and they are subject to change from release to release as Mathworks continues to try to improve the algorithms, so there is no set value of "N" intact rows that can be stated will ensure the correct interpretation.
The best and most reliable choice when reading a number of files of a given structure is to create and save an import object that defines precisely how you want the data to be interpreted; then it won't matter where or how many missing data are there. You can also control what is done with missing data although not with total granularity. Note that the import object does not have to be created each time for each file; having one that describes the specific data format and using it for all files is the most efficient method. While you can control a lot of the options by using the optional named-parameter pairs, specifying the data type by variable/column is not one that can be done except by the import options object.
