> No, not to the eieio--class but to the new object.

Right, I was confused.  I left the word FIXME but rephrased the comment
so that we don't mention the non-existent dflt.  At the moment, I'm not
ready to reimplement this with aset and to ensure it would not break by
accident in an obscure manner.

> But could you add a test or two to
> test/lisp/emacs-lisp/eieio-tests/eieio-tests.el ?

I added sort of exhaustive tests for initialization.  A complete
exhaustive test would also take :default-initargs and inheritance into
account but I'd rather do this gradually.  I did not run tests in Emacs
28 but they pass as is in Emacs 27.

Some :initform's needed a fix (quote), some could be improved (search
for “symbol-value” in the patch).

The necessity to quote should be expected to break some packages, maybe
a lot.  E.g. helm contained one unquoted such instance.  I already fixed
it; nothing else broke for me so far but it was easy to omit this quote.

If authors get confused about initform workings, it may help to note
that :initform actually expects a *form*, i.e. expression to be
evaluated.  The attempt to guess what to evaluate and what not was based
on the wrong premise that CL does that.