Note: the ULTIMATE geo god (who is so pro of a geo god that I don't feel privileged to refer to him by his aops username) points out that the same polynomial argument can be carried out in $\mathbb{Z}_{p^k}$, with "lagrange-style" polynomials
$f(x): \mathbb{Z}_{p^k} \to \mathbb{Z} = \sum\limits_{j=0}^{p^k-1} f(j) \binom{x-j-1}{p^k-1}$, which works because $\binom{x-j-1}{p^k-1}$ is divisible by $p$ unless $x-j-1\equiv p^k-1 (\bmod\; p^k)$ which means $x=j$ and $\binom{x-j-1}{p^k-1} \equiv 1(\bmod\; p)$.
CANBANKAN wrote:
Keep a running counter $c$ the starts at $p-1$ and decreases to $0$. Let $l_j$ be the $x^c$-coefficient of $P_j$. By inductive hypothesis on $k-1$, Ali can get all $l_j$ to be 0 if they play with this sequence. Reflect a move with $b_j=z$ with moves with picking $b$ st $b_{j+tp^{k-1}}$ forming a polynomial with leading term $zx^c$. Now decrement c by 1. This eventually guarantees that $a_{j+tp^{k-1}}$ does not depend on $t$ so we apply inductive hypothesis where $b_j$ is uniquely defined by j mod $p^{k-1}$.
To clarify, we are playing the $(p^{k-1},p)$ game on $l_0, l_1, \cdots, l_{p^{k-1}-1}$. Suppose in the $(p^{k-1},p)$-game strategy, Ali chooses the sequence $b_0, \cdots, b_{p^{k-1}-1}$. Then Ali will choose polynomials $X_j$ for $0\le j<p^{k-1}$ where the leading coeff of $X_j$ is $b_jx^c$. Then if Mohammed chooses to rotate them by $tp^{k-1}+r$, $l_j$ becomes $l_j+b_{j+tp^{k-1}+r}=l_j+b_{j+r}$ (if I rotate by $p^{k-1}$, the $b_j$'s aren't affected as the $X_j(t)$ become $X_j(t+1)$, whose leading coefficients don't change.) This justifies it being the same game as $(p^{k-1},p)$ because I have $l_0,\cdots,l_{p^{k-1}-1}$ and need to make them 0.
He will make all the $P_j$'s have degree at most $p-1$, then $p-2, \cdots,$ and eventually $0$, which implies $a_j=a_{j+p^{k-1}}=\cdots=a_{j+(p-1)p^{k-1}} \forall j$. Now A adapts a strategy on $(a_0,\cdots,a_{p^{k-1}-1})$ and chooses $b_j=b_{j+p^{k-1}}=\cdots=b_{j+(p-1)p^{k-1}}$ in remaining moves, so Mohammed's shifts are equivalent to shifts mod $p^{k-1}$, so a strategy for $(p^{k-1},p)$ finishes.