What people should care about is the integrity and consistency of the data they're storing. RDBMS do an extremely good job of generalising this problem with reasonable performance, which suits the majority of real world problems. The ideas behind it are understood by most of the industry.

It should be the default choice (or plain filesystem storage), unless you have a specific requirement for something different. In the latter case, this should be people who are well informed about the database choices they're making.

Unpredictable performance would mean, you can't predict how long a specific query will take. This is the opposite of how postgres works. If a query is slow, it's predictably slow if the data doesn't change. So what you must mean is, when the data changes, it doesn't behave like you expect.

So again, how do these poor choosers make better decisions in the face of changing data when they have to write all the algorithms by hand, and understand how to scale the data correctly? Seems to me that pretty soon they'd just be writing a very crappy database on top of some keystore.

The only way for query performance to be truly predictable is for your data to be static. If that's not your situation then a query optimizer is incredibly helpful because it means that you don't have to rewrite your queries every time you add an index.