IMO to make socionics practically testable, one would have to create an experiment that puts concrete data (numbers, results) to the notion of compatibility between people.
The next thing would be to propose a method of establishing typings that is does not in any way avail itself of subjective asessments. Ideally you'd want a method that, with a little help, a contemporary computer could employ. Again, this would involve feeding the computer concrete data gathered under experimental conditions.
Put the two together, and you have a scientifically testable claim: does the method of establishing typings predict the results of the experiments or does it not.