Fix incorrect data returned in a corner case of partial-select HRI
While investigating better reporting on disagreeing metadata and actual
results returned by the data source, I stumbled across an incorrect
optimization applied to the HRI fast-path in
aa1d8a87. The reason this has
never been a problem in the wild is that the failure case is very convoluted:
In order for the problem to present itself one needs to have a subchain of
... something-single ... with the "something" not fetching anything
AND the entire chain being hit exactly once (no multiplication of the branch
neither by a result or a parallel 1:M) AND the HRI fast-path code needs to be
in effect. Then and only then everything from "something"onwards will present
as "nonexisting left join" and a sole undef will be returned in place of the
entire substructure.