The attached patch series is modified to improve estimates for these special clauses (Var op Var with the same var on both sides) without extended statistics. This is done in 0001, and it seems fairly simple and cheap.
The 0002 part is still the same patch as on 2021/07/20. Part 0003 fixes handling of those clauses so that we don't treat them as simple, but it does that by tweaking statext_is_compatible_clause(), as suggested by Dean. It does work, although it's a bit more invasive than simply checking the shape of clause in statext_mcv_clauselist_selectivity.
I do have results for the randomly generated queries, and this does improve the situation a lot - pretty much all the queries with (a=a) or (a<a) clauses had terrible estimates, and this fixes that.
That being said, I'm still not sure if this is an issue in real-world applications, or whether we're solving something because of synthetic queries generated by the randomized generator. But the checks seem fairly cheap, so maybe it doesn't matter too much.
