[step:Fix the rank-selection convention and reduce to one deleted rank]
Let $\rho:P\to \{0,1,\dots,r\}$ be the given rank function, and let $\hat{0}$ and $\hat{1}$ denote the minimum and maximum elements of $P$. For $S\subseteq \{0,1,\dots,r\}$, define the rank-selected subposet
\begin{align*}
P_S=\{x\in P:\rho(x)\in S\},
\end{align*}
with the order inherited from $P$. When $P_S$ is nonempty, its rank function is the order-preserving relabelling of the selected ranks in increasing order.
It is enough to prove the following one-rank deletion statement:
[claim:Deleting one rank preserves Cohen-Macaulayness]
Let $Q$ be a finite bounded Cohen-Macaulay graded poset over $k$ with minimum element $\hat{0}_Q$, maximum element $\hat{1}_Q$, and rank function $\rho_Q:Q\to \{0,1,\dots,m\}$. If $j\in \{1,\dots,m-1\}$ and
\begin{align*}
T=\{0,1,\dots,m\}\setminus\{j\},
\end{align*}
then the rank-selected subposet
\begin{align*}
Q_T=\{x\in Q:\rho_Q(x)\in T\},
\end{align*}
with the inherited order and the relabelled grading on the selected ranks, is Cohen-Macaulay over $k$ in the finite graded poset convention: its whole order complex and the order complexes of all its open intervals are Cohen-Macaulay over $k$, with the void-complex convention in dimension $-1$.
[/claim]
[proof]
The proof occupies the next steps.
[/proof]
Indeed, given $S\subseteq \{0,1,\dots,r\}$, define the omitted interior-rank set
\begin{align*}
J_{\mathrm{int}}=(\{1,\dots,r-1\}\setminus S).
\end{align*}
Starting from $P$, delete the ranks in $J_{\mathrm{int}}$ one at a time. After some interior-rank deletions, the remaining original ranks form a subset $R\subseteq \{0,1,\dots,r\}$ that still contains $0$ and $r$, so the intermediate rank-selected poset is still bounded, with inherited minimum $\hat{0}$ and maximum $\hat{1}$. Its induced rank function is the order-preserving bijection from $R$ to $\{0,1,\dots,|R|-1\}$. To delete the next omitted original rank $j\in J_{\mathrm{int}}\cap R$, apply the one-rank deletion statement to the relabelled rank corresponding to $j$ under this bijection. This relabelled rank is neither the minimum rank nor the maximum rank, so the next intermediate poset remains bounded. Since $P$ is Cohen-Macaulay over $k$, induction gives a bounded Cohen-Macaulay rank-selected subposet $P_R$ whose original rank set is
\begin{align*}
R=S\cup(\{0,r\}\setminus S).
\end{align*}
It remains only to remove endpoint ranks if $0\notin S$ or $r\notin S$. If both endpoint ranks remain in $S$, then $P_R=P_S$. If one or both endpoint ranks are omitted, then $P_S$ is obtained from the bounded Cohen-Macaulay poset $P_R$ by taking either an upper open interval, a lower open interval, or the proper part $(\hat{0},\hat{1})_{P_R}$. The whole order complex of that unbounded endpoint-deleted poset is therefore one of the interval complexes already included in the Cohen-Macaulay condition for $P_R$, and its open intervals are also open intervals in $P_R$. Hence $P_S$ is Cohen-Macaulay over $k$ under the finite graded poset convention stated in the theorem.
If $S=\varnothing$, then $P_S$ is the empty poset. Its order complex is the dimension-$-1$ complex specified in the theorem statement, and it has no nonempty open intervals. If $S$ has one element, the whole order complex of $P_S$ is a finite discrete pure $0$-dimensional complex and every open interval in $P_S$ is empty, so the Cohen-Macaulay condition holds by the same dimension-$-1$ convention for the interval complexes.
[/step]