When we run this test, we create a lot of workspace watches, and these are very expensive at the database. This makes it hard to isolate scaling issues with the actual autostart process we are trying to test.
A related problem is: coder/coder#21337 which makes the watches expensive.