Hmm, when I join all my eight incremental database schema changes into just a single one (basically drop support for migration of old databases), my test execution time drops from about 1:10 minutes to just 33 seconds. I might consider doing exactly that. I’m the only one who runs that software anyway.
Just haven’t figured out where exactly the speedup comes from. I suspected that the column recreation is kind of expensive, but it doesn’t really appear to that obvious. More testing is needed.