This won't break existing code. It will break new, good code that gets back-ported from eg. a 3.7 tutorial to a 3.5 environment, without any syntax errors.
> Changed in version 3.7: LIFO order is now guaranteed. In prior versions, `popitem()` would return an arbitrary key/value pair. [1]
If they added a new `popordereditem()` method, good 3.7 code would use that, and an attempt to run that code on 3.5 would throw a reasonable, useful error message. If they wanted to play it safe like Rust or Go, they'd add the ordered method and make popitem() deprecated or make it artificially use random/arbitrary order so you can't accidentally depend on a new implementation detail and have a test case work.
Also, your case 4 does deserve some protection because while bad code it's hard to test for bad but working code. Implementation-specific or undefined behavior that works is the worst kind of problem, because it's hardest to test against. Compile-time syntax errors are the easiest, runtime errors are testable with a solid test harness, some possible can be identified as warnings with linters, but sloppy code requires manual inspection to detect.
Actually, no, I take that back: Implementation-specific or undefined behavior that works sometimes! is the worst kind of problem. That's what this code enables; if your test case is `dict({'one': True, 'two': True})` and you have a unit test where you popitem() to assert that you get 'two' and then 'one' it will pass the test harness on 3.7, it will pass on 3.6 because of implementation-specific behavior, and it will pass on 3.5 because the hash map arbitrarily puts those particular items in order. But it will silently get it wrong when you pass in user-supplied data. Shudder.
[1]: https://docs.python.org/3/library/stdtypes.html#dict.popitem