When determining if a dependent join should be used we'll rule out too many scenarios based upon requiring additional ndv info - when that can be determined from cardinality alone. Also we are inappropriately assuming key usage in predicate costing - as it does not consider the node which may not represent a single table.