There is a serious but trivial bug in the copy constructor for NormalVariableImpl.  It does not initialize m_nextValid, causing junk m_next results to be returned.  Beyond being an incorrect distribution then, the results may not obey the bounds.  This came up for me in that NormalVariable was still occasionally creating negative values, causing the scheduler to bail.

Note that m_nextValid should be initialized to false, rather than copying the value, as m_next is not copied (which sounds reasonable to me).

The fix for this is trivial, patch attached.

Thx