• The feature was a mess. It seems to work as intended on perl 5.13.8 and newer. In older version, the optimizer incorrectly rejected some matches
• the intended behavior is that a missing no-pattern is the same as an empty no-pattern, ie the whole match succeeds if the condition fails and the no-pattern is missing. Thus both regexes should succeed.

The condition (?=b) must return false.

It does. The condition returns false, therefore it make no attempt to match the (null) yes pattern (which would have matched anyway), and simply moves on to the next element 'a$', which matches.

In other words, if the condition of the conditional expression, in this case the lookahead assertion, is false, then no attempt is made to use either the yes pattern, or the no pattern (here missing). In effect, the entire construct becomes a noop.

and simply moves on to the next element.

And that's the bug, IMHO. If the yes-pattern is present, and the condition fails, the whole match normally fails. Why should it move on if the match fails? Compare:

'a' =~ /^xa/ # match of x fails, doesn't move on to match a
[download]

When the no-pattern is not present, and the condition fails, it normally doesn't match:

$ perl -wE 'say "a" =~ /^(?(?=b)xxx)a/'

$ # no match
[download]

So it doesn't show the "move on"-behavior when the yes-pattern is something else than the empty pattern.

So it's a special case. And it's not documented. So it's a bug.

Perl 6 - second systems done right

Why should it move on if the match fails?

Is "(?(condition)yes-pattern)" an abbreviation for "(?(condition)yes-pattern|)" (i.e. with an explicitly empty no-pattern)? This seems most reasonable to me, but I find nothing in the documentation to suggest whether this or any other possibile behaviour should be expected.

The most fundamental "bug" seems to be lack of documentation of intended behaviour.

"In effect, the entire construct becomes a noop. "

In which case the first line of Serge314's example should match, shouldn't it?