>> Jim, can you also try the patch and report if you see regressions?
>
> Thanks. I tried this out for a bit, and I haven't been able to break it 
> yet (although I didn't notice the issue that spawned this bug# the last 
> time either, so it's possible I'm missing some problem this time around, 
> too).
>

Thanks for your confirmation.

Stefan, do you agree with the attached patch?  Or would you like something 
else?

Eli, are you okay with that patch?

The difference between the code before dafa6d6bad and the code with this 
patch applied can be summarized as follows:

  (defun pcomplete-arg (&optional index offset)
    (let ((arg
           (nth (+ (pcase index
  	           ('first 0)
  	           ('last  pcomplete-last)
  	           (_      (- pcomplete-index (or index 0))))
  	         (or offset 0))
                pcomplete-args)))
-    arg
+    (if (or (stringp arg)
+            (eq index 'last))
+        arg
+      (propertize
+       (car (split-string (pcomplete-actual-arg index offset)))
+       'pcomplete-arg-value arg))))

IOW, instead of unconditionally returning arg, we now sometimes (when arg 
is not already a string and when index is not last) return the textual 
representation of that argument (what the user typed in) with the value of 
arg attached to it.