> This makes an obviously visible difference: C-x 5 2 opens a small frame
 > which immediately becomes large.
 >
 > Here's the *frame-size-history* buffer:
...
 > ConfigureNotify, PS=1328x1260, XS=400x374, DS=1328x1260
 > xg_frame_resized, rejected, PS=1328x1260, XS=400x374, DS=664x630
 > tool-bar-lines (2), MS=160x175
 > xg_frame_set_char_size, visible, PS=1328x1260, XS=1328x1260, DS=1328x1260
 > ConfigureNotify, PS=1328x1260, XS=1328x1258, DS=1328x1260
 > xg_frame_resized, rejected, PS=1328x1260, XS=1328x1258, DS=664x671
 > set_window_configuration (4), MS=160x175 IH IV
 >
 > Note that once more I retained the simple patch to frame.c to set
 > frame-size-history to (100).

Setting 'frame-size-history' to '(100), doing C-x 5 2 and evaluating
'frame--size-history' should be enough.  The frame.c change is needed
only for the initial frame because at that time no Lisp was evaluated.

The result is not yet what I hoped for.  We reject two ConfigureNotify
events but the ensuing xg_frame_set_char_size calls do not get us a
ConfigureNotify we'd accept.  IIUC the final height may be off by two
pixels.  What does (frame-text-height) return for the first and the
second frame?  Also I hope you didn't get the GTK assertion failure.
And how does removing the menubar behave?

Now something completely different.  A couple of years ago I noticed
that our size hints calculations are fundamentally wrong.  I tried to
fix them and I attach some parts of that fix together with the previous
changes I proposed.  Attached as size_hints.diff.  If the patch does not
apply, complain.  I edited it by hand to exclude other changes I made
and I might have made mistakes.  Again I'm interested in the history of
the second frame.

martin