You bring up a good point, but I think the takeaway is that you need to be thoughtful about the process, not necessarily that one way or the other is always right.

How often are breaking changes introduced to the APIs of OpenBLAS or LAPACK? I would guess approximately never (I could be wrong though. I'm actually surprised and a bit disturbed how much activity the GitHub repos are still getting).

Beyond that, how much of the API do you need, or are likely to need in the future? Why not implement that matrix multiplication yourself to start, but abstracted in a way where you can drop in OpenBLAS in the future? You can even copy their API if you don't want to abstract, and include a bit of profiling to warn you if it suddenly gets very slow on a certain platform.

In my experience the chances are good you'll never need that last 20%.