Chris Kuklewicz
The authors of sed are in agreement with your intuition. But I think policy 2 as a recursive definition it is unusual. And I see policy 2 as very hard to summarize.
The single-match is always easy to summarize: the leftmost longest match.
Policy 1 for multiple matches can be summarized as:
All leftmost longest non-overlapping matches, where overlapping means sharing the same matching characters.
Policy 2 for multiple matches can be summarized as:
All leftmost longest non-overlapping matches, where overlapping means sharing the same matching characters, excluding zero-length matches that coincide with the end of a non-zero-length match.
Policy 0 has been:
All leftmost longest non-overlapping matches, where overlapping means sharing the same matching characters, stopping with the first zero-length match.
Given this point, policy 1 does indeed seem the most consistent behaviour. Thanks for the explanation!