Maybe auto-update should verify with the user whether it wants to update an entry that has manually edited fields? I'm thinking of the use case where a lot of the references are marked for auto-update in a batch. I guess the work-around is to de-select every element known to be hand-curated, but this seems prone to error. The issue is that hand-curation of a reference takes a while, so it is a bit of a loss if that work is overwritten without any way to restore it or preempt it.
I can see how crowd-sourced versions of the metadata might go the wrong way, but a halfway solution might be to let users see different bibtex options sourced from the canonical sources and from other users, and then pick their favorite? Scholar already lists multiple versions of papers, often with unique bibtex files (most of them rubbish, supporting your argument about crowdsourcing...)
