Speed and cost. DuckDB runs on your machine, starts instantly and needs no cluster, so a golden test takes seconds and spends no warehouse compute. That makes it practical to run the tests on every save, not only before a release.
DuckDB is the judge for the expected outputs; the same transformation is then translated into your platform's dialect for deployment. DuckDB cannot show everything, though: permissions, performance on full data and platform-specific behavior only appear on the platform.
That is why each template's cases also run on the real platform before the template is marked ready there, and why each platform profile lists known traps that lint rules check for. A template is only called ready on a platform after it passes on that platform. See golden tests on DuckDB.