We basically ended up creating a tool+language spec that would let us define the mapping of input to output sequences. We wrote it in such a way where we sat with scientists and pretty much mapped all the possible cases that they could think of for valid inputs and what the code should produce.
Then this tool would basically write automated tests for us, in such a way as to not only test correct behavior, but also do combinations of inputs out of order, fuzz testing, and so on. We ended up making it also check memory state, to ensure that there was no memory leaks, and analyze the memory space for required data or data that should not be there.
In the end, whenever someone was developing anything for this software, they would basically just run the tests, and it would be very good at catching possible errors, mostly on the negative side (i.e for a fuzzed input, it would result in an output that should be an error).
We could have done the same thing with a typed language, but it would have to be very strict typing to the point of something like CoQ, and it would have taken us probably the same amount of time to write that.
I think you're arguing that tests and types can both be used to check a particular case for correctness. Sure, this is true. However "moving type checking code into the test suite" means nothing—that would just be type checking.
When you make a code change, there's a difference in the kinds of feedback you get from your tests and your types. Tests usually cover business logic or stories—things that you want or don't want. Types ideally cover everything. They check every operation applied to a given piece of data. Of course types are rarely precise enough that they can catch every logic bug (e.g. strings with semantics not encoded in types, like email addresses).
This is just scratching the surface. You might just have to try out both to get a better sense of how they feel to work with.