Documentation, bug report, mailing list discussions, something like that. In particular I'm interested in the questions:
* Why are they required in combination with failing statements (when every other database does an "automatic savepoint and rollback" for a failed statement)? * When was that behavior chosen? Was it always like that? Was it adopted later? * What is the design decision behind it?
There is a lot of information on what savepoints are and how they work (and also thanks to you I'm now fairly certain I have good grasp on them), but I fail to locate documentation on these questions.
Those questions would not be answered in user-facing documentation. You can explore the git history and search the far-back mailing list archives if you wish to satisfy your curiosity. For me this is how it works - the only question for me is whether I should argue that the behavior should be changed. I do vaguely recall this topic coming up in the recent (couple of years) past...but changing transaction behavior like this is problematic no matter how desirable the new state might be to have (and that's debatable).
It may be worth updating the docs here but you have received your official answer - I'm nearly positive I'm right and even if I was mistaken most likely I would have been corrected by now. I am writing this on a mailing list...
