3D mesh refinement gives strange mesh sizes

[Marie Rognes]

Something slightly peculiar seems to happen with 3d refinement. Refining
a 1 x 1 x 1 Unit Cube 3 times gives largest cell diameter that is
increasing (but smallest cell diameter that is decreasing as expected).
It does not seem to (only) be a bug in hmax()/hmin(), odd convergence
rates are also observed.
Example:

from dolfin import *

mesh = UnitCube(1, 1, 1)

for i in range(3):
    mesh = refine(mesh)

    print "Refinement level = ", i
    print "h_max = ", mesh.hmax()
    print "h_min = ", mesh.hmin()

gives:

Refinement level = 0
h_max = 1.65831239518
h_min = 0.866025403784
Refinement level = 1
h_max = 1.47901994577
h_min = 0.433012701892
Refinement level = 2
h_max = 1.7809758561
h_min = 0.216506350946

[Anders Logg (logg)]

Should be fixed now:

hmin = 1.73205 hmax = 1.73205
hmin = 0.866025 hmax = 0.866025
hmin = 0.433013 hmax = 0.433013
hmin = 0.216506 hmax = 0.216506
hmin = 0.108253 hmax = 0.108253
hmin = 0.0541266 hmax = 0.0541266
hmin = 0.0270633 hmax = 0.0270633
hmin = 0.0135316 hmax = 0.0135316

The problem was a missing check for how to partition the middle octahedron to avoid increasingly bad aspect ratios.