> What you're arguing here is that because there are some cases where pure Python is simply not fast enough, there is no point in thinking about performance at all in a Python program. It's a false dichotomy.

No, it's a very real dichotomy. If your complexities are right (e.g., you're not using a quadratic algorithm where you could be using a linearithmic algorithm) and Python still isn't fast enough for your task, switching from a hash table to a binary tree is not going to make it fast enough.