In fairness to Postgres, “DO NOTHING” is accurate within the SQL domain if you squint. It’s reasonable for queries to have side-effects on the underlying implementation as long as the invariants defined by SQL are met. Surprising, yes, but the alternative would be leaking implementation details in the syntax.

On the other hand, it’s so easy to use and rely on autoincrement IDs and to assume they are monotonic and predictable based on naive testing. If we could do it all again I’d fuzz the IDs even during normal operation (occasionally increment by a random extra amount) so that they don’t become such a foot gun for developers.