Unfortunately, this isn't the right way to fix the issue. We already have infrastructure in place that creates synthetic symbols for index-signature-based property and element accesses (see getApplicableIndexSymbol), but it isn't working right and it also is odd in that it creates symbols whose type you can't obtain using getTypeOfSymbol. I will be putting up a PR that fixes this.

is this still needed after #3018?

This is kind of a pure port from Strada - you can see how a bunch of failingTests are removed here. This branch is already based now on top of #3018 - so the diff you see is not outdated or anything like that.

Hmm, as I said above, I don't think this code is right--even if it was in Strada. Some of those tests change quick info on an error location from displaying nothing to displaying any (which is actually the errorType). How is that better?

As a user, that displayed any is what I expect - it might be an acquired expectation but now it would feel slightly off to me if those wouldn't display anything.

Yes, I would tend to agree, that if we are hovering on an identifier-y thing, we need to display something...

I gotta say I disagree here. If the type returned from GetTypeAtLocation is the error type (meaning there's no type), how are we helping anyone by displaying any? Which is indistinguishable from an actual non-error any.

To some extent, it's about cross-location consistency:

function test(o: unknown) {
    //   ^? function test(o: unknown): any
    return o.foo
    //       ^? any
}