In the documentation for the supported arithmetic operations for a datetime object, there is the following note, among other:

datetime2 = datetime1 - timedelta

Comment:
Computes the datetime2 such that datetime2 + timedelta == datetime1. As for addition, the result has the same tzinfo attribute as the input datetime, and no time zone adjustments are done even if the input is aware. This isn’t quite equivalent to datetime1 + (-timedelta), because -timedelta in isolation can overflow in cases where datetime1 - timedelta does not.

While reading the source code for __sub__ operation I found in the first few lines:

Line 1885:
def __sub__(self, other):
    "Subtract two datetimes, or a datetime and a timedelta."
    if not isinstance(other, datetime):
        if isinstance(other, timedelta):
            return self + -other
        return NotImplemented

Is the documentation in contradiction with the actual implementation?

The C "_datetime" implementation seems to handle this as documented. But either way, the "timedelta" range is greater than the "datetime" range, so it seems to be just a difference in OverflowError messages, not a big practical problem:

Python "datetime" implementation:
>>> import sys
>>> sys.modules["_datetime"] = None
>>> from datetime import *
>>> datetime.max - timedelta.max
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\Martin\AppData\Local\Programs\Python\Python35-32\lib\datetime.py", line 1741, in __sub__
    return self + -other
  File "C:\Users\Martin\AppData\Local\Programs\Python\Python35-32\lib\datetime.py", line 518, in __neg__
    -self._microseconds)
  File "C:\Users\Martin\AppData\Local\Programs\Python\Python35-32\lib\datetime.py", line 430, in __new__
    raise OverflowError("timedelta # of days is too large: %d" % d)
OverflowError: timedelta # of days is too large: -1000000000

C "_datetime" implementation:
>>> from datetime import *
>>> datetime.max - timedelta.max
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
OverflowError: date value out of range

I agree.  The documentation can be improved here.  The note about x - y not being quite the same as x + (-y) belongs to the timedelta - timedelta operation. It should be removed from both date - timedelta and datetime-timedelta footnotes.

New changeset 5b6e49a1393b3e2313471696e3568e26296137b4 by Alexander Belopolsky (Farhaan Bukhsh) in branch 'master':
bpo-30516: Fix documentation issue with -timedelta in datetime (GH-7348)
https://github.com/python/cpython/commit/5b6e49a1393b3e2313471696e3568e26296137b4

New changeset 55f39bdabc91b387c18451c78ee077e6d05d6cfe by Alexander Belopolsky (Miss Islington (bot)) in branch '3.6':
bpo-30516: Fix documentation issue with -timedelta in datetime (GH-7348) (GH-8092)
https://github.com/python/cpython/commit/55f39bdabc91b387c18451c78ee077e6d05d6cfe

New changeset a8bb18bbb9c286cfac5a0b8c8679c440e5c49305 by Alexander Belopolsky (Miss Islington (bot)) in branch '3.7':
bpo-30516: Fix documentation issue with -timedelta in datetime (GH-7348) (GH-8093)
https://github.com/python/cpython/commit/a8bb18bbb9c286cfac5a0b8c8679c440e5c49305