This change introduces a serious slowdown which is noticeable for large

candidate lists (try with 10000 elements.) The slowdown happens on every
invocation.

It is obvious that having duplicate candidates makes no sense, but at
the same time scanning the list in advance for all duplicates is
expensive.