We recently received a letter from ECMWF dated July, 1991.  Here is 
the relevant text:

It has recently been brought to our attention that, on a regular lat/long
grid, the ECMWF u- and v- components of wind are incorrect at the poles.
Currently, the horizontal components of wind give inconsistent polar values
for wind magnitude.  The problem affects users who

(i)   have received...data from the ECMWF/TOGA basic data sets for periods
      before 1 July 1991 [NCAR dataset ds111.2];
(ii)  [deleted, applies only to ECMWF Advanced sets.  At NCAR, ds111.0 is
      in spherical harmonic representation, and ds111.1 and ds111.3 are on 
      a Gaussian grid, not a lat/lon grid.]
(iii) have received...data from the ECMWF/WMO Analysis data sets [NCAR 
      dataset ds110.0].

These users can derive consistent values for the horizontal components of
polar winds using the nearest neighbouring latitude circle.

As from July 1991 changes will be made to the interpolation routines used to
create the ECMWF/TOGA Basic data sets....  As a result of these changes,
the data will have the following format:

(a)  Surface data:  the grid points at each of the poles will contain
     horizontal wind components from the nearest neighbouring Gaussian 
     latitude circle interpolated to the required resolution.  For the T106
     model, the nearest latitude circle is +/-89.14 degrees.  With the 
     introduction of the new T213 model, which is expected to be in
     September 1991, it will be +/-89.578132 degrees.  [NCAR is still 
     receiving T106 data, to our knowledge.]

(b)  Upper Air data:  the grid points at each of the poles will contain the
     correct horizontal wind components, i.e. the values of wind magnitude
     derived from the horizontal wind components will be constant, while the
     u and v components oscillate with a wavenumber 1 pattern around the
     pole.