I guess using _ variables was valid before, so this is a feature and a breaking change, instead of a bugfix?

No, they weren't valid before. We were just missing not assigning to them in the codegen.

So it's a bugfix? I see.

Indeed :-)

At some point "_" was a valid variable and block argument name. Eventually we made "_" be a variable that can be assigned to but never read. This is represented with the Underscore AST node. But then "_" was still allowed as block argument name. Changing that to use an Underscore AST node was very hard, so I just check against a "_" name. I implemented this but forgot to add the line above. It worked with just one "_" block argument though: that variable was assigned but could never be read. With two "_" that variable could be assigned twice with different types and that's where the bug happens. This PR changes the codegen so that "_" is never assigned a value.