I noticed one thing here, that did not change: if e.g. mixed[] is unpacked and pushed onto the empty array then it would evaluate to a truthy value as in https://phpstan.org/r/91ebaf38-ad93-499f-8f84-6382b13df4a8.
This is not correct and should not error IMO, as the array could be empty, then nothing is pushed and we have a falsy value. I wanted to improve this here too, but failed because of the constant arrays that are involved. They always lead to true/false values when calling toBoolean().

Please push a test with https://phpstan.org/r/91ebaf38-ad93-499f-8f84-6382b13df4a8, I have an idea about it :)

Pushed new tests and the implementation of the idea that I had as well :) It is a bit more complicated than I hoped it would be, maybe you have a better one.
Update: the longer I think about it, the more I can come up with weird edge cases or simplification ideas. I'll take another look tomorrow.

OK, I ended up basically re-writing the array_push / array_unshift handling. Not sure if that's ok, sorry 😅

It seems like this is quite well covered via both LegacyNodeScopeResolverTest - binary.php and NodeScopeResolverTest - constant-array-optional-set.php. At least those were the ones that started failing while I was working on this.
Maybe it is also possible to strip out even more common stuff from the 2 blocks that I currently have (constant array handling, other handling). done

found another thing (#1225) that will make things fail here atm, marking as draft for now again..

From my POV this would be ready. This PR improves array_push and array_unshift in general now. IMO it still simplifies it a bit, but tbh I don't like the current complexity much either. Parts of it are because of different handling for ConstantArrayType but I don't think this can be handled better atm.

The 10 failing phpstan-symfony checks should be unrelated.

Looks complicated, but probably worth it :) Thanks.