Every time I've worked on—or seen other people work on—a "we must hurry at all costs project", enough time is wasted on debugging, firefighting and friction that just building a decently clean system from the get-go would have been faster. I've seen this unambiguously on "throw away" data science code meant to produce a single report (or normal science code meant for a single paper), and that's a much shorter-term effort than anything at a startup. The tradeoff is even clearer once you start working in weeks and months, much less quarters or years.