For a high-level paradigm, we need to formulate a model specification
paradigm.


Lessons from lattice:

 * Formulas and non-standard evaluation has drawbacks, especially for
   non-top-level calls, wrappers, etc.  However, we do still need to
   think about enclosing environments.

 * A data-centric approach is good, but the data frame model is too
   limited.  In other words, the design should anticipate other data
   containers (we have a ready-made use-case in flowSets)

 * Distinction between primary and conditioning variables is good, but
   giving special status to 'x', 'y', etc. has drawbacks.  A simple
   use-case is densityplot(x=, weights=)

 * The ability to specify multiple data sources is useful (something
   lattice cannot do well, but ggplot can).





Ideas:

 * Specify variables through expression-like objects, with eval-like
   functionality, but make the tools explicitly generic to work with
   non-standard data sources.  Think about attaching an enclosing
   environment to these expression-like objects.

 * Make the idea of packets concrete.  Perhaps a cross-tabulation-like
   array, each element an index vector (subscripts, but potentially
   more flexible).  A "packets" object should have an attached data
   object (could be .GlobalEnv).  packet[[i]] should have enough
   information to extract the corresponding packet from a different
   packets object (with the same cross-tabulation)

 * Variable evaluation API:

     evaluate(expression, data, subset)
     evaluate(expression, packet[[i]])


 * High-level function calls could be of the form

     mosaiq(data = environment(), enclos = .GlobalEnv,
            margin.vars = list(a = expression(a), b = expression(b)),
            packets = packets(margin.vars, data),

            panel.vars = list(x = expression(x), weights = expression(w)),  # specific to panel function
            panel = qv.panel.densityplot,
            prepanel = qv.panel.densityplot,

            ...)



 * One convenient convention is to make each panel function accept an
   argument called 'give.limits', so that when called as

     panel(..., give.limits = TRUE)

   it essentially works as a prepanel function.



 * Outline of steps and related functions


    - compute.packets : computes packets as a list array 

    - compute.layout : computes layout --> 
        will need to match these to packet order; 
	generates integer-array of dim c(column, row, page), with
	packet number as entry (0 for no packet)

    - compute.limits : call prepanel function for all packets --> 
        generates similar list-array, for now of the form list(xlim, ylim)

    - combine.limits : combine limits information across packets ->
        (1) support types same, free, sliced
	(2) allow choice of margins over which to combine 
        --> results as a list-array again

    - create.panels : create list-array of qwidgets, with structure
	similar to layout

    - create similar list-array of strip.[top|left],
	xaxis.[left|right], yaxis.[top|bottom]

    - render into page widgets, one for each page in layout[,,page]:
	+ draw all panels
	+ draw strip.top for row=1:nrow (just 1 for at top)
	+ draw strip.left for column=integer(0) (just 1 for left)
	+ etc. for axes