0.14.7

• A floating-point warning inside an expression is a warning again. An expression's inline
  arithmetic (x1 + x2, x1 * x2) runs in its own frame, so numpy raises its floating-point
  warnings there (inf - inf, an overflow in a product), and Python's warning machinery looks
  for __builtins__ in that frame's globals. The evaluation namespace did not carry it, so such
  an evaluation raised KeyError('__builtins__') instead of returning nan with a
  RuntimeWarning whenever numpy's error policy was warn (its default). The namespace now
  binds __builtins__ explicitly. Bare names resolved through the interpreter's builtins before
  and still do, so what an expression can reach is unchanged.