This is a good point! We've gotten some instances of Chesterson's Fence that some devs casually remove during a "test refactor" that later allowed a regression to make it into prod.
I've caught some errors in "test refactoring" from our multiple levels of testing having large overlap with teach other. Our end-to-end tests have a lot of overlap with the integration tests which in turn have a large overlap with unit testing. The unit tests run in a matter of seconds compared to the end-to-end which can take minutes or in some cases hours for our manual testing so the levels of testing also serve as an efficiency for us.
I've caught some errors in "test refactoring" from our multiple levels of testing having large overlap with teach other. Our end-to-end tests have a lot of overlap with the integration tests which in turn have a large overlap with unit testing. The unit tests run in a matter of seconds compared to the end-to-end which can take minutes or in some cases hours for our manual testing so the levels of testing also serve as an efficiency for us.