>   . the FIXNUMP assertion is probably there for a reason; what is that
>     reason?

Before 572fe798cd0a00ad4a9050a7962cf8e8fbcc209b (from 2014-09-30), the computation of left and top was done in w32_createwindow:

      /* When called with RES_TYPE_NUMBER, w32_get_arg will return zero
        for anything that is not a number and is not Qunbound.  */
      left = x_get_arg (dpyinfo, Qnil, Qleft, "left", "Left", RES_TYPE_NUMBER);
      top = x_get_arg (dpyinfo, Qnil, Qtop, "top", "Top", RES_TYPE_NUMBER);

and anything not a number was turned to 0.

In that commit you moved the code to my_create_window and used XINT:

 +  /* When called with RES_TYPE_NUMBER, x_get_arg will return zero for
 +     anything that is not a number and is not Qunbound.  */
 +  left = x_get_arg (dpyinfo, Qnil, Qleft, "left", "Left", RES_TYPE_NUMBER);
 +  top = x_get_arg (dpyinfo, Qnil, Qtop, "top", "Top", RES_TYPE_NUMBER);
 +  if (EQ (left, Qunbound))
 +    coords[0] = CW_USEDEFAULT;
 +  else
 +    coords[0] = XINT (left);
 +  if (EQ (top, Qunbound))
 +    coords[1] = CW_USEDEFAULT;
 +  else
 +    coords[1] = XINT (top);
 +

and these XINTs were transformed into XFIXNUM (which easserts) as part of a big XFIXNUM/XFIXNAT change  (by Tom Tromey, in commit d1ec3a0a8e4d7d56ebc1e4fa743130b9974ac6a8 from 2018-08-07).

So, from my ignorance, it seems like the idea was always to set non-nums to zero, and the assertion came by accident.

    Juanma