Automodel: Switch tests to inline expectations

[max-schaefer]

This entails factoring the where clauses of our queries into predicates we can use in the tests, but I think that's a good idea anyway since it allows us to get the full results of the queries without sampling if desired, which can be quite useful when comparing results from multiple runs (since the sampling makes diffs very noisy).

I also cleaned up the treatment of unexploitable types (primitives, boxed types, and void). Previously, some candidates with such types were not suppressed in framework mode, and in both modes we had two overlapping characteristics for excluding them, leading to duplicate results for the negative-examples queries.

Consequently, on apache/geode the number or negative examples has gone down a bit (519 -> 430 in application mode, 6215 -> 6146 in framework mode), and so has the number of candidates in framework mode (40326 -> 39065).

[kaeluka]

Thanks for a very nice PR.

I've left some inline comments, but no blocker.

This entails factoring the where clauses of our queries into predicates we can use in the tests, but I think that's a good idea anyway since it allows us to get the full results of the queries without sampling if desired,

💯 yes, as we have noticed, the sampling used in some extraction queries makes the testing very tedious and also error-prone (eg., when a candidate goes away it's not obvious whether the query logic removed it, or the sampling did).

FYI, @tausbn: you will be pleased to hear we'll finally be using inline expectation tests (I think you created an issue sometime to do exactly that, but I can't find it 😬)

[tausbn]

FYI, @tausbn: you will be pleased to hear we'll finally be using inline expectation tests (I think you created an issue sometime to do exactly that, but I can't find it 😬)

I think maybe I never wrote up an issue. I just groused about it a lot on various PRs. 🙂

Anyway, yay! 🎉

[kaeluka]

Looks great!

I also restarted a failing test (looked flaky)