SQL Server can produce incorrect results when reading from a partitioned non-unique nonclustered index with a descending leading key.
The provided script demonstrates the issue with a simple seek, but the bug is quite general. It caused incorrect results from a Merge Join in a production application, for example. The sort order required by the Merge Join was not delivered by the index.
The bug appeared in SQL Server 2008 and persists through SQL Server 2014 CTP 1.
SQL Server 2005's APPLY model partitioning correctly handles the scenario.