I don't really think accidents of implementation are different from hard requirements in Python, as it applies to alternative implementations. In practice if it deviates from CPython then it's broken. There is no language spec, there is only CPython. This has been the experience and approach of PyPy as far as I've understood it after having followed their blog over the years.
Definitely not true. There have been times when other implementors have come to python-dev and said, hey, is this part of the spec or is it an implementation detail? And the answer determines whether they care about that or not. For just a few examples:
1) Reference counting vs nondeterministic garbage collection 2) O(1) indexing/slicing of Unicode strings 3) "\N{...}" string escapes (okay, that's standardized, but optional) 4) Reuse of id() values 5) The "slots" mechanism for dunder method lookup
The language spec determines, in some cases, that a CPython implementation detail has been promoted to standard. More often, it determines that other Pythons are permitted to behave differently.
Sometimes they will come away from this list thinking they don't care but then their users will report bugs over and over again and they'll just have to do it anyway. You probably won't hear about most of those. Trees that fall in the forest when no one is there do in fact make a sound. / Anders