If a project passes the tests in Ruby 2.0.0-p648 and Ruby 2.3.1, does it make sense to also test with versions such as 2.1.8 and 2.2.3?
Has there been any language features that worked in Ruby 2.0 and Ruby 2.3 but were temporarily not working or worked differently in for example Ruby 2.2?
You should test your code for environments you intend it to be used in. Ruby language version is one thing that can vary. You might also consider testing against JRuby and Rubinius if your code is supposed to support it - e.g. it is provided as a public gem.
Logically, testing on earliest and latest versions should cover most failure scenarios with respect to language features (although not necessarily all language bugs since new ones can be introduced). As far as I know there has not been a Ruby feature that was deliberately added or removed in one version and then that decision reversed in a later version. Unless: Perhaps in your production code you are detecting a feature's existence and then using the feature in full - in which case an intermediate Ruby version which has the feature but not in its latest state could fail.
There may be other caveats too, and philosophically speaking when you start testing you want to avoid too much logical "this should work because . . ." thinking. The point of testing is to demonstrate that your code doesn't fail in ways you have covered (well, there's more depth to it than that, but the answer would get far too long if it dove into test philosophies). If you want to declare "works in all versions of Ruby MRI from 2.0.0 to 2.3.1" then you will feel safer making that statement if you have actually tested it. In fact when making such a statement in a public place, I would usually just say something closer to raw fact - "tested in versions 2.0.0, 2.1.4 and 2.3.1".
Obviously there are diminishing returns. If you have no problem in 2.1.9, it is very unlikely you will have a problem in 2.1.10 - at some point it will cost you more to check every minor variation, even just to look at the test results, than the benefit of extra coverage.
The usual answer to this problem is to test as many variations as your automated test environment can handle and you can be bothered to set up and maintain. If using multiple versions of Ruby is done for you in parallel by your test service provider - e.g. you are using Travis - then it is relatively cheap to test multiple versions. So you may as well get as much coverage on environment variations that you expect to see "in the wild" as you can be bothered to look at.
