NaNs with mixed quadrature degree and coffee

[FabioLuporini]

So this one is turning out a little tricky to reproduce. But with the following code:

from firedrake import *
#parameters['coffee'] = {}
mesh = UnitSquareMesh(1,1)
V = FunctionSpace(mesh, "RT", 2)
P1 = FunctionSpace(mesh, "CG", 1)
v = TestFunction(V)
u = Function(V)
c1 = Function(P1)
c1.assign(0.001)
c2 = Constant(1e-6)
dt = Constant(1.0)
dx2 = dx(degree=5)
F = inner(v,u)*dx - dt*(- c2*inner(v,u)*dx2 - c1*(-u[1]*v[0] + u[0]*v[1])*dx2)
print assemble(derivative(F,u)).M.values.sum()

I'm getting NaNs in the assembled matrix (most of the time!) if I leave COFFEE optimisations switched on. I say most of the times, as running the script multiple times I sometimes get different answers (sometimes the correct one), so I expect some sort of initialised value bug? It seems to be also very sensitive to the precise form of the equation.