mapcan/mapcon (new test)

Functions mapcan and mapcon have two edge cases where most implementations make mistakes. (Including ccl, abcl, ecl, cmucl and clisp; sbcl just fixed that on the mainstream; clasp handles it right.)

Per http://www.lispworks.com/documentation/HyperSpec/Body/f_mapc_.htm (mapcon f x1 ... xn) should be equivalent to (apply #'nconc (maplist f x1 ... xn)) and similarly (mapcan f x1 ... xn) to (apply #'nconc (mapcar f x1 ... xn)).

1. All but last arguments to nconc must be proper lists, so (nconc 1 2 3) should error. (mapcan #'identity '(1 2 3)) and (mapcon #'car '(1 2 3)) are equivalent to (nconc 1 2 3), so they should error too, but the mentioned implementations return 3 instead.

2. nconcing lists with shared structure is valid (there is an example in specification: http://www.lispworks.com/documentation/HyperSpec/Body/f_nconc.htm), so it should be valid to do in mapcan and mapcon as well. Here is an example where the mentioned implementations go into an infinite loop:

(setf *print-circle* t)

;; This returns #1=(1 2 . #1#)
(let ((lst (list 1 2)))
  (flet ((f (x)
           (declare (ignore x))
           lst))
    (apply #'nconc (mapcar #'f '(1 2)))))

;; Should be equivalent to the previous example, but hangs instead.
(let ((lst (list 1 2)))
  (flet ((f (x)
           (declare (ignore x))
           lst))
    (mapcan #'f '(1 2))))