Is it always possible to diagonalize a Clifford gate $ g $ using a gate $ V $ from the third level $\mathcal{C}^{(3)}$ of the Clifford hierarchy? In other words can every Clifford gate be written as $ g=VDV^{-1} $ where $ V $ is from $\mathcal{C}^{(3)}$ and $ D $ is diagonal?
Update: Just rehashing the argument from @AdamZalcman mostly for my own benefit. Any matrix which can be expressed $ VDV^{-1} $ for $ D $ diagonal and $ V \in \mathcal{C}^{(3)} $ is of the form $ aI+bVZV^{-1} $. When $ b=0 $ then all scalar matrices can be obtained this way. When $ a=0 $ then all $ VZV^{-1} $ for $ V \in \mathcal{C}^{(3)} $ can be obtained this way. $ VZV^{-1} $ has projective order 2 (squares to a scalar matrix) and it is in the Clifford group (since $ V \in \mathcal{C}^{(3)} $ and $ Z $ is Pauli) so it must be either a non identity Pauli or a Hadamard type matrix, up to global phase. The final case to consider is $ a,b $ both nonzero. The only Clifford gates that can be written as a nontrivial combination of $ 1 $ and a Pauli are the square roots of Pauli gates. However no Clifford gates can be written as nontrivial combination of $ 1 $ and a Hadamard (such a matrix would have Pauli coordinates exactly three of which are nonzero, no Clifford gate has such coordinates see second column in table below). Thus we have exhausted all Clifford gates which can be written as $ aI+bVZV^{-1} $ for $ V \in \mathcal{C}^{(3)} $ and we finally conclude that no order 3 single qubit Clifford gate can be written as $ VDV^{-1} $ for $ V \in \mathcal{C}^{(3)} $.
Note that throughout the argument we an remove the up to global phase/ up to scalar ambiguity by specializing to the case of determinant 1 Cliffords. Then this shows determinant 1 Clifford gates with (projective) order 3 cannot be written as $ VDV^{-1} $ for $ D $ diagonal and $ V \in \mathcal{C}^{(3)} $.