What's the specific breaking change this will expose ... if it's behavior that people would never want, and no tests caught it, then I could imagine us not taking this as a major change -- worst case perhaps we could put this behavior behind a flag, to help people bumping into issues, and then swap the default value in the next major.

Thinking on this one a bit, a couple things jump out at me:

1. Rather than implicitly deciding whether to concat values, based on whether a default value has been configured for a key, better that we figure out a way to track (during parsing) whether a default value has been used -- we've talked in the past about adding an isDefaulted method, this would be a great feature for users, and would also allow us to simplify this logic.
2. Providing config feels different than default values, and I could imagine 50% of people wishing that values would be combined, and the other 50% wishing for the opposite behavior. This makes me think that rather than taking this as a breaking change, we'd do better to introduce a parser flag, to allow folks to configure which approach is taken.