I also noted some other things about the D200 RAW data looking at the
light bulb files.  

The clipping value is only 4095 in every other column of pixels; in the
remainder, they alternate at something like 4000 and 4016, which means
that any value above 4000 really should be clipped, as clipping at
different levels creates false detail in the highlights.  I didn't think
to make a correlation between this and the shadow banding, but if the
issue is an offset, they are probably both symptoms of the same problem.

The Canon 10D does this variable-clipping point also, in a more
random-looking pattern, but the stretch is only over about 9 levels
(3997 - 4006 at ISO 100, with the range sliding up at higher ISOs).  RAW
converters should ideally clip all this data at the lowest clipping
value (after optimally blackpointing, of course).

The RAW data is already blackpointed in the NEF.  This allow more
potential headroom, but as seen above, it is mostly wasted, as compared
to a camera that outputs RAW data to 4095, with a positive blackpoint.
I believe that keeping the positive blackpoint offsets in the data makes
it easier for future converters to eliminate noise at the RAW source.
On the 20D, where the blackpoint is generally 128 in the RAW data,
binning the RAW data for downsampling can bring the average noise close
to the linear blackpoint (128); binning already-black-clipped data does
not make the average black level approach linear black, as all of the
binned elements are 0 or greater.

The DNG 3.3 converter apparently not only drops the masked, black pixels
from D200 NEFs (maybe the NEFs don't even contain them, as they are
useless if the data is already blackpointed), but the DNG converter
crops the exposed RAW image by 4 pixels horizontally (3900 instead of
3904, as output by IRIS).