[CF-metadata] bounds

[Brian Eaton]

Hi Jonathan,

I agree with specifying the ordering of the 2nd index for 1D bounds.  I
propose that we require data writers to follow the constraint
bnd(i,1)=bnd(i+1,0) exactly, rather than to allow for tolerances.  It's
analogous to the requirement that the exact value of _FillValue must be
written to variables.  The netCDF User's Guide doesn't allow for tolerances
when identifying the _FillValue.

I also agree with specifying the order for the 2D cases where a single
cell vertex index is used.

As I review the discussion about bounds for 2D retangular grids it seems
that the motivation for having a special case is to be able to treat the
common case of contiguous grids as efficiently as possible.  Steve has
suggested the n*m*2*2 encoding.  Bob has stated that he sees no advantage
to the n*m*2*2 encoding over the n*m*4 encoding.  I'm guessing that Steve
sees an advantage to n*m*2*2 because once the index ordering is specified
one can work out which vertices are shared by adjacent cells and quickly
determine that the grid is contiguous.  In the n*m*4 case, even if the
order is specified, the starting vertex isn't and so making the
determination of contiguity is not quite as easy.  In any case, most of the
discussion has been about how to determine whether or not a grid is
contiguous.

If the special case of interest is actually contiguous 2D rectangular grids
then what advantage does the n*m*2*2 encoding have over (n+1)*(m+1)?  The
reasons for not allowing the (n+1) encoding of 1D coordinate boundary data
don't apply to this special case.

Brian