Continuations and constructor-hoisting

#lang racket

(define xs #f)
(set! xs '()) ;ensure Racket will permit rebinding of xs at the REPL
(define cont #f)

(define (meh x)
  (let ((y (vector x 10)))
    ;(display y)
    (if (test x)
        y
        x)))

(define (test x)
  (call-with-current-continuation
   (lambda (c)
     (set! cont c)))
  (even? x))

(set! xs (cons (meh 8) xs))

(cont 'ignored)

(eq? (car xs) (cadr xs))
;=> #f, which I believe is wrong.

;If you uncomment the (display y), then this returns #t.
;[Also, Chicken returns #t regardless.]
;The uncommenting appears to *force* the compiler to create the
;vector y before (test x), and so any continuation saved from
;there will retain the identical y.

;On the other hand, if Racket accepts this as wrong, this makes
;it impossible for it to hoist constructors into if-branches
;unless it can prove that the test doesn't leak any continuations
;or it does other tricky things.
;There are some places in the standard that say "it is an error to
;have multiple returns from this special form", but I don't
;think "if" is one of them.