If the goal is to exclude json-like constructs only, then I think we should try a little harder to weed out illegal json. Some ideas: can we remove variable references, function literals, regexp literals, calls?

(We have previously tried to classify entire json-like files as generated, but we never found a scalable way to to that reliably.)

Looking great! 👍

We should probably exclude CodeInAttribute top-levels from this check, otherwise LGTM.

Some ideas: can we remove variable references, function literals, regexp literals, calls?

I've removed function literals from the check for now. When I look at the results we no longer flag I only want to remove functions from those.
The rest with e.g. binary expressions look like a benign parser-tests or something.

We should probably exclude CodeInAttribute top-levels from this check, otherwise LGTM.

I don't think there is any need for that.
Code in an attribute is already not in a void-context, and thus something like <div onclick="[1,2,3]"></div> is already not flagged.

An evaluation looks fine.
Lots of parser tests are no longer flagged as useless.