Enhanced test suite

[Rich-Harris]

Describe the problem

We occasionally regress in ways that aren't covered by the test suite, because of things like dev vs prod, HMR vs not.

Describe the proposed solution

It would be impractical to run every possible combination on every test run — there are thousands of tests, and the test suite already takes a decent chunk of time to run — but maybe we should do it on changesets PRs so that even if a PR with a regression gets merged, it doesn't get released.

We would only need to do this on one Node/OS combination.

Importance

nice to have