This run using actions/cache@v2 shows how a cache created by the update-deps job is not hit by the jobs in the test matrix that depend on update-deps and use the same key. This works in actions/cache@v1, so it looks like a regression to me.

What's more peculiar, each of the three test jobs reported to have successfully stored its cache state under this same key.
In the next run, all jobs hit the cache, but which state did they get?