FWIW, the promised semantics of saturating arithmetic require that _keep_positive be run on entire the entire counter:

    >>> c1 = Counter(a=-3, b=4)
    >>> +c1
    Counter({'b': 4})

    >>> from collections import Counter
    >>> c1 = Counter(a=-3, b=4)
    >>> c2 = Counter(b=5, c=5)
    >>> c1 += c2                 # The "a" entry gets removed
    >>> c1
    Counter({'b': 9, 'c': 5})

When this behavior isn't wanted, use the update() method which is documented not to perform removal of non-positive values.  That method is fast in pure python and has an even faster C helper function:

    >>> c1 = Counter(a=-3, b=4)
    >>> c2 = Counter(b=5, c=5)
    >>> c1.update(c2)
    >>> c1
    Counter({'b': 9, 'c': 5, 'a': -3})

And as mentioned before, the Counter API is not encapsulated -- the underlying structure is fully exposed, so a user is free to do anything with it that they can do with a regular dictionary.

A final thought is to be careful when proposing to rewrite the internals of the existing methods.  These methods are careful to retain ordering of inputs, to not use auxiliary data fields (it is just a dict variant with no other structure), and to make only minimal assumptions about the datatype for the count.  They are designed to not be magical.