My last job was perforce (no unsubmitted commit stacking) with everyone developing directly against a single branch. In theory, code reviews were supposed to happen before hand.

In practice, I submitted small changes from the hip, and complained when my coworkers gave me a week long spitball of unrelated features in a single mega-commit. Not that I wasn't guilty of that myself. Bleh.

    - def f(x: X): Option[X]
    + def f(x: X): MyErrorType \/ X

Some changes simply cannot be broken into tiny atomic pieces. Often these are important changes for long term maintainability.

The Right Way (TM) to make that change, then, is to make only the change to the result type in the first commit (along with the pattern matching necessary to ignore it), followed by a commit for each place that the updated result type can be usefully used.

Change the code without changing the behavior: i.e. REFACTORING. Run the tests, verify you didn't break anything, then start writing tests for the behavior you wish to change, followed by changing that behavior.

If you think you're at atoms you probably haven't seen strings.

If you fix the behavior everyplace the compiler complains, you are guaranteed to resolve most issues (ignoring overloaded methods, e.g. getOrElse). If you use a toOption method on the union type to make a tiny atomic change, the compiler will no longer complain in all the places that mistakes are being made.

I suppose an intermediate way might be to rename f to f_DEPRECATED, replace everywhere, and build the new method returning the right type. But in my experience its better just to bite the bullet and change everything.