Sed makes one choice. I'd guess that GP would call this broken garbage too, and I'd agree. Regular expressions have all these nice theoretical properties like closure under all the boolean operations and linear-time matching, but these nice properties get trashed by features that don't mesh or aren't fully thought through.
In this case (thinking about capturing groups and commutativity), one property of regular expressions is that for each one there is a machine that can do the linear-time matching -- a DFA. Even if the regular expression contains not-mutually-exclusive alternations, when it gets compiled to a DFA, the matching procedure is deterministic by construction. I can imagine a way to integrate capturing start and end actions into the transition edges of the DFA. The right thing to do is to perform capturing on all the matching alternands, not just the first. You lose the ability to number the capturing groups left to right, but instead you should lay them out in a tree that follows the concatenation/alternation structure of the expression.
Just write /(a)b/ or /a(b)/ or /ab/ or /(ab)/ or /(a)(b)/ which mean five slightly different things.
It is not reasonable to expect the user to manually disambiguate every regex.
Assuming input is "ab",
Probably the easiest way to test this yourself is with GNU sed.