This was an interesting bug.  Let's first define a recipe to make it more 
apparent:

C-x C-m C-\ french-postfix RET C-x ( pa^te'_a`_l'Unicode RET C-x ) C-x e

This should insert "pâté_à_l'Unicode".  It worked as expected with Emacs 
21-24.

After commit 30a6b1f814, which was the result of the discussion in 
http://lists.gnu.org/archive/html/emacs-devel/2015-08/msg00193.html , and 
before commit 03e3440dbb, it inserts "paâtteé__aà__l'UUnniiccocododee". 
This is what one gets in Emacs 25 and 26.  This is not surprising, as the 
recipe in that discussion and quail.el both use unread-command-events, but 
expect an opposite effect.

After commit 03e3440dbb, which was the result of bug#32108, it inserts 
"pâtté__à__l'Unnicocdode".  This is what one gets in Emacs 27 (and 28 till 
now).

I attach a patch to fix that bug.  I checked that the recipes that led to 
the two above commits still work correctly, and that it does not introduce 
regressions with make check.

Note that the inhibit--record-char variable, which was introduced by 
commit 03e3440dbb, is, after applying that patch, only used once, namely 
in lisp/term/xterm.el, as a result of bug#44908.  It is not used in ELPA 
or MELPA.  I'm not convinced that bug#44908 is a sufficient reason to keep 
that variable.