bug#41423: 27.0.91; eshell file completion in tramp dir is slow (3 minutes) [regression on pretest]

[Gregory Heytings via "Bug reports for GNU Emacs, the Swiss army knife of text editors" <bug-gnu-emacs@gnu.org>]

Apparently my previous last note was not the last one ;-)

I still don't know how this bug should be fixed (except by using (setq 
completion-at-point-functions '(pcomplete t))), but here is a more 
detailed explanation of what is happening, at least how I understand it:

1. TAB calls completion-at-point
2. completion-at-point calls pcomplete-completions-at-point, which calls pcomplete/cd; this completes the directory name
3. completion-at-point let-binds completion-in-region-mode-predicate to a lambda, which contains pcomplete-completion-at-point
4. completion-at-point calls completion-in-region, which adds completion-in-region--postch to post-command-hook
5. post-command-hook calls completion-in-region--postch
6. completion-in-region--postch funcalls completion-in-region-mode-predicate
7. this calls pcomplete-completions-at-point a second time, which again calls pcomplete/cd (and adds a '/' after the directory name (?))
8. RET is pressed
9. post-command-hook still contains completion-in-region--postch: it is called again
10. completion-in-region--postch funcalls completion-in-region-mode-predicate again
11. this calls pcomplete-completions-at-point a third (!) time
12. at this point pcomplete-completions-at-point considers (for some reason, possibly because the last character of the input is '/' (?)) that it is a command that it must now complete
13. therefore instead of calling pcomplete/cd a third time, pcomplete-completions-at-point now calls eshell-complete-commands-list
14. this loops through all possible command names (on the local or remote host)
15. if all command names had had a common prefix, that prefix would have been inserted now, but this is not the case, so the effect of this loop (apart from a waste of time) is nil

The fact that default-directory is remote is not important here, the exact 
same steps take place when it is local, except that step 15 is executed 
much faster.

At step 15 it is possible to interrupt the loop with C-g.  At step 8 it is 
possible to remove completion-in-region--postch from post-command-hook for 
example by switching buffers with C-x C-b.

Why such a overly complicated mechanism is used, or what should be done to 
avoid this behavior, is beyond my understanding.