I think there's a lot of valuable stuff here. At the same time I don't like the need to add cast() in various places.

I wonder if this could be done in smaller steps, e.g. make the changes to checker.py and checkexpr.py without the other changes? They should still work with Expression and Statement just aliases, and you'd get a fair amount of the "more readable code". Then you can iterate on the rest, possibly implementing selftype first.

I don't like the cast either. But how does keeping the alias - effectively, cast() everywhere - better than a couple of explicit casts? There's no content to the Expression/Statement classes (which makes me think that a huge Union might be a better choice. is it?)

If the motivation for making only incremental change is affecting less lines of code at once, it's fine with me.

Looks like you can get rid of three casts by dropping the distinction between LvalueExpr and Expression. The union there doesn't sit well with me either, so maybe you can just get rid of of LvalueExpr for now (I know you have plans there, but please be patient).

One reason I dislike casts so much is that they incur real runtime overhead (it's a plain Python function and nothing optimizes it away).

Another reason is that we've had several occurrences where a cast was wrong (the 2nd argument was not an instance of the 1st argument) that were hard to debug.

I'd still like to see if we can do this more gradually -- I hate huge diffs, they are a pain to review and often get bogged down by (or cause) merge conflicts. So maybe one file at a time, with the complicated stuff last, if the tests can be made to pass in between?

I agree that casts must be obviously-correct, I will try to break this change to smaller pieces.

Turns out the Declaration thing does not hold water. I think there's some deep mix with Var, Argument, and SymbolNode-stuff. It should be sorted out, but not in this PR.

I have a suggestion. Maybe we can should start by doing the easy part only -- change most occurrences of Node to either Expression or Statement, but keeping those as aliases to Node for now. That will buy us the documentation benefit without having to worry about the subtler parts of the diff. We can then quickly approve the part of the diff that causes massive churn, and even if we end up not doing the other part ever it'll still be useful (if only as a hint). And the actual fix to make everything type-check correctly will be a much smaller diff which will make it easier to review (believe it or not) and less likely to go stale due to a merge conflict.

If you agree, you can choose to close this and start a new PR, or remove the tricky bits from this PR and resubmit it for review. Hope this helps!

Closing. See above.