I ran the dump_plot script with the change and when I noticed nothing failing, I pushed it here:
http://hg.savannah.gnu.org/hgweb/octave/rev/24e11ee4fe45

Marking as ready for test.

There is probably a better way to deal with the 3d extents of 2d objects that would solve this more generally. I'll try to think about it and report back.

Definitely an improvement.  I would say that this resloves the problem with 2-D aspect ratios.  But a problem still remains with 3-D aspect ratios.  Take the same code, but display the third dimension.

t = linspace (0, 2*pi, 100);
y = sin (t);
plot (t, y);
axis equal
view (3)
## This line causes plot to disappear
pbaspect ([1500 750 1])

If I check with the Z limits with zlim() I find [-1, 1].  The plot can be made to appear again if I set the Z limits to something smaller such as [-.002 +.002].  This appears to be what Matlab does.  If I set the limits with

zlim ([-1, 1])

in Matlab then the plot disappears there as well.

I think Matlab will adjust the limits of the data, as long as "zlimmode" is "auto", when "dataaspectratio" and "plotboxaspectratio" are adjusted.  My guess is that Octave does not do that.  Maybe there is another special exception going on, because it does update the X and Y limits when I enter the command

pbaspect ([1500 750 1])

The axis limits in z-direction are not calculated correctly:

>> pbaspect ([1500 750 750])

>> get (gca, "zlim")
ans =

  -1.0708   2.0708

>> pbaspect ([1500 750 1])
>> get (gca, "zlim")
ans =

   0.49791   0.50209

In the second case (of the disappearing plot) the z-limits don't include 0.
Matlab sets these limits symmetrically to z=0 for 2d plots which seems to make more sense.