I'm having a problem with the limits in the presence of missing
values.  The two choices I have in the current design are
unappealing:

1.  Determine the limits including the missings when they're
    shown.   This will cause the graphics to scale in a reasonable
    way, but the limits shown in the table won't be the ones we want
    to see.

2.  Determine the limits excluding the missings at all times.
    (This is what's happening now.)  The limits in the table look
    right, but the graphics need to be rescaled to get the missings
    into the plot.

Other options:

Do I somehow have to distinguish between scaling limits and
display limits?  How does that interact with the fact that I
already have five (columnwise) sets of limits: 
  lim_raw, lim_tform, lim_specified, lim_specified_tform, lim

I've always been keeping track of the overall missing limits
in d->lim_missing.

I'd better review the use of all these limits before doing
anything.

** The limits that are displayed in the table are definitely
the transformed limits, and I think that makes sense.
** The user-specified limits, though, are raw.

I just added lim_display, which is the limit that will show up
in the table.  Now lim_tform will conform to choice 1 above,
and lim_display (together with the mean and median) to choice 2.

There are still oddities in the way missings are handled.  If
missings are represented as 0's, for instance, those are treated as
real data values for some purposes but not for others.  If the
variable is transformed and the missings are displayed, the 0's will
also be transformed; that's probably not very sophisticated.
