Is there a clear rationale documented some where for why adding new attributes to a metric is a breaking change?

This PR comment is one of the earliest assertions that I'm aware of that adding attributes is breaking, but it has remained unclear to me whether folks have a shared understanding of why.

ADDITIVE changes to metrics can break user alerts by creating new timeseries and adjusting thresholds. I think we need to be a lot more careful around changes, particularly around metrics here.

I do not personally have experience with a backend where this is true, but if it is true for some backend, I think it would be useful to articulate this some where.