Let $X(t)= \phi X_{t-12}+Z_t+\theta Z_{t-1}$ where $Z_t\sim WN(0,1)$. I need to find prediction error for projecting $X_t$ onto $H_{t-3}(X)$ (Hilbert space).
So, I know that $X_t \perp P_{H_{t-3}}X_t$ what gives me:
$\|X_t\|^2-\|P_{H_{t-3}}X_t\|^2=\langle \phi X_{t-12}+Z_t+\theta Z_{t-1}, \phi X_{t-12}+Z_t+\theta Z_{t-1}\rangle-\|P_{H_{t-3}}X_t\|^2=\sigma^2+\theta^2\sigma^2-\|P_{H_{t-3}}X_t\|^2$
but what to do with $\|P_{H_{t-3}}X_t\|^2$?
Remark: Mind, that we can solve this only is $(X_t)$ is casual. A sufficient condition for that is $|\varphi|<1$. Using the representation with a lag operator $B$ \begin{equation} (1-\varphi B^{12})X_t = (1 + \theta B)Z_t. \end{equation} We can evaluate the casual form of $(X_t)$, which is \begin{equation} X_t = (1+\theta B)\sum_{i = 0}^{\infty}\varphi^iX_{t-i}. \end{equation} The sum does not converge when $|\varphi| \geq 1$.
Remark: You said, you knew that $X_t\;\bot\;P_{H_{t-3}}X_t$, which is not true in general. The truth is $X_t - P_{H_{t-3}}X_t\;\bot\;P_{H_{t-3}}X_t$ and it gives as: \begin{equation} \text{Err}(P_{H_{t-3}}X_t) = ||X_t - P_{H_{t-3}}X_t||^2 = ||X_t||^2 - ||P_{H_{t-3}}X_t||^2. \end{equation} The Pythagorean theorem "applies" here with $a = X_t - P_{H_{t-3}}X_t$, $b = P_{H_{t-3}}X_t$ and $c = X_t$, in its form $a^2 = c^2 - b^2$.
As we have casuality the problem goes much more easily - have a look at $H_{t-3}$ set. It contains only those elements that can be expressed as linear combinations of $X_{t-3}, X_{t-4} \ldots X_{t-12} \ldots $. Because of causality $Z_{t}, Z_{t-1}, Z_{t-2}$ are perpendicular to $H_{t-3}$, so the forecast is just \begin{equation} P_{H_{t-3}}X_t = \varphi X_{t-12}. \end{equation} You can imagine $X_{t-12}$ as only one element that the projection operator $P_{H_{t-3}}$ do not bring to zero (as it does with perpendicular elements). It works as in usual $\mathbb{R}^3$ space with perpendicular vectors to a projective space.
Another usefull fact is that the covariance function of $(X_t)$ is: \begin{equation} \gamma_X(0) = \frac{\sigma^2(1 + \theta^2)}{1 - \varphi^2}. \end{equation}
Also we have $\mathbb{E}X_t = m \in \mathbb{R}$ and: \begin{equation} \begin{split} ||X_t||^2 & = \langle X_t - m + m,\, X_t - m + m \rangle \\ & = \langle X_t - m,\,X_t - m \rangle + 2\langle X_t - m, \,m \rangle + \langle m,\,m\rangle \\ & = \gamma_X(0) + 2\mathbb{E}[(X_t - m)m] + m^2 \\ & = \gamma_X(0) + 2m\mathbb{E}X_t - 2m^2 + m^2 \\ & = \gamma_X(0) + m^2. \end{split} \end{equation} But we can assume that $m = 0$ when we are considering projections. Simply center all variables in $(X_t)$ sequence, make a projection and shift them back by $m$, leaving an error untouched.
At this point, our problem is very simple: \begin{equation} \begin{split} \text{Err}(P_{H_{t-3}}X_t) & = ||X_t||^2 - ||P_{H_{t-3}}X_t||^2 = \gamma_X(0) - ||\varphi X_{t-12}||^2 = \gamma_X(0) - |\varphi|^2 \gamma_X(0) \\ & = (1 - \varphi^2) \gamma_X(0) = (1 - \varphi^2) \frac{\sigma^2(1 + \theta^2)}{1 - \varphi^2} = \sigma^2(1 + \theta^2). \end{split} \end{equation}
