= Do something about words without vowels? (what about abbreviations?)
= Do pairs of words, not just singles.
= Hash by soundex -- probably won't work out very well.
= Randomly bounce to a similar sounding word -- sometimes, or all the time,
  or weighted by frequencies of the respective words.
= Make it possible to load in a saved file, then suck more data into it
  (make files.c be able to create either `words' or `pwords'.)
= Make it able to descend directory trees.
= Make it able to talk HTTP?  NNTP?
= Do interesting typography: enumerated lists, subtitles, (Do it with HTML,
  too.)
= Insert random whitespace to make it look like poetry.
= Count syllables to make haiku.  Loop regenerating sentences until we get
  ones that have word and sentence breaks in the right places.  Counting 
  syllables is hard -- have to snarf a hyphenation algorithm from somewhere.


Here are some ideas I plan to someday implement for v2.0, unless someone
beats me to it:

  * pwords don't point to strings, because they aren't just single words
    any more.  the key is the string, and the pword object itself contains
    only statistics about the string under which it was indexed.

  * pword->succ always points to pwords that are single words.
    but the pwords table contains both kinds.

  * to record text:
    for each sentence
      there is an N-word buffer, initially empty
      for each word
	for each cdr of the buffer (0-N, 1-N, 2-N, ... N)
	  look up the cdr in the pwords table
	  index this word under the pword we just found
	add this word to end buffer, dropping the old word 0 off the end

  * to generate text:
    while generating sentences
      pick a random sentence-start word
      there is an N-word buffer, initially empty (which is <= value of N)
      while not done
        pick a random number R, from 0 to min(N, size-of-buffer)
        look up a pword in the table, using the last R words as the key
        (while there is no match, --R and try again)
        emit that word
        push that word onto the end of the buffer

The idea here is that, given N=3, and the input text

	All work and no play makes Jack a dull boy

when scanning that text we would record these associations:

	"all"		  --> "work"

	"all work"	  --> "and"
	"work"		  --> "and"

	"all work and"	  --> "no"
	"work and"	  --> "no"
	"and"		  --> "no"

	"work and no"	  --> "play"
	"and no"	  --> "play"
	"no"		  --> "play"

	"and no play"	  --> "makes"
	"no play"	  --> "makes"
	"play"		  --> "makes"

	"no play makes"	  --> "jack"
	"play makes"	  --> "jack"
	"makes"		  --> "jack"

	"play makes jack" --> "a"
	"makes jack"	  --> "a"
	"jack"		  --> "a"

	"makes jack a"	  --> "dull"
	"jack a"	  --> "dull"
	"a"		  --> "dull"

	"jack a dull"	  --> "boy"
	"a dull"	  --> "boy"
	"dull"		  --> "boy"

When generating:

  * suppose we already have picked the sequence "all work and no".
  * pick a random number from 1-N.
    * if 1: look up "no"
    * if 2: look up "and no"
    * if 3: look up "work and no"

  * pick a random successor to the pword associated with the string we
    looked up ("play".)

  * now we've got "all work and no play".
  * repeat.

If "work and no" wasn't found in the table (meaning we generated it, but
it never occurred in nature) then we would decrease N and look up "and no".
If that didn't match either, we'd look up "no".  That way, word sequences
which had actually occurred together are more likely to be chosen than
ones that didn't.

The N used for generating can't usefully be larger than the N used for
recording.  The N used for recording is an important tunable paramter,
that is probably language-centric; I'd guess that 3 is good for English.
The larger N, the more likely one is to regenerate the exact input text.