Some of the worst examples of that are when a project uses a custom build of a library of which the source no longer exists and no record of the changes exists either.
If there are changes in future numpy versions we want, it's up to us to backport them, which is nowhere near our core business.
There's a lot to be said for standardization and 'boring.'
Now add in a script that updates this artefact with the latest version, breaking changes, and it all goes wrong and it's hard to find the correct previous version of the binary artefact to build your code any more. Especially if the build has been broken for a while due to the project being on the back-burner and it quietly dies when no-one is looking.