I think the answer to your question is in this paragraph of the docs.

Empty leading (or trailing) fields are produced when there are positive width matches at the beginning (or end) of the string; a zero-width match at the beginning (or end) of the string does not produce an empty field.

I say I think, because, even with the example given, I find it quite difficult to interpret that sentence. Especially in the light of a couple of other sentances

By default, empty leading fields are preserved, and empty trailing ones are deleted.

and

If LIMIT is unspecified or zero, trailing null fields are stripped

and

Note that splitting an EXPR that evaluates to the empty string always returns the empty list, regardless of the LIMIT specified.

Not at all clear to me:)