First of all, note that since $n > 2$ we must have $x \ne y$ since equality cannot be achieved if $x = y$. Now let $x, y$ be distinct primes and $n > 2$ be an even integer such that the equation holds.
<insert proof that $x = 2$ here; my previous proof was flawed>
... therefore $x = 2$. Now observe that:
$$x^n + x^{n - 1} + \cdots + 1 = \frac{x^{n + 1} - 1}{x - 1}$$
Therefore:
$$2^n + 2^{n - 1} + \cdots + 1 = 2^{n + 1} - 1$$
So that we are now solving the following equation for $y$ an odd prime (recall $x \ne y$) and $n > 2$ integer:
$$2^{n + 1} - 1 = y^2 + y + 1 ~ ~ ~ \iff ~ ~ ~ y^2 + y + \left ( 2 - 2^{n + 1} \right ) = 0$$
Using the quadratic equation discarding the negative solution we find:
$$y = \frac{-1 + \sqrt{1 - 4 \left ( 2 - 2^{n + 1} \right )}}{2} = \frac{\sqrt{2^{n + 3} - 7} - 1}{2}$$
We are therefore now solving the diophantine equation in $n > 2$ and $z$ given by:
$$2^{n + 3} - 7 = z^2$$
This is a variant of the Ramanujan-Nagell equation given by $2^m - 7 = x^2$, and its solutions are in one to one correspondence with the solutions to our equation, by a direct change of variables $(m, x) \to (n + 3, z)$. It is known that the Ramanujan-Nagell equation has only five solutions in $m$, given by $m = 3, 4, 5, 7, 15$. These translate to $n = 0, 1, 2, 4, 12$, of which only the solutions $n = 4$ and $n = 12$ are admissible. Working backwards, this gives:
$$n = 4 ~ ~ ~ \implies ~ ~ ~ z = 11 ~ ~ ~ \implies ~ ~ ~ y = 5$$
$$n = 12 ~ ~ ~ \implies ~ ~ ~ z = 181 ~ ~ ~ \implies ~ ~ ~ y = 90$$
But $y = 90$ isn't prime, and so does not satisfy the conditions, and we therefore conclude that the only solution $(x, y, n)$ is $(2, 5, 4)$.
Okay, so using Ramanujan-Nagell was a bit of a cop-out, I guess. I have a proof of it in front of me and it's not what I would call elementary, and I'd be interested to see solutions that don't need to invoke it, especially since I didn't manage to prove that there are no other solutions with $x \ne 2$; perhaps the restriction to $y$ being a prime is essential to solving the challenge efficiently... I also just noted that it's easy to show that $n$ must divide $y - 1$, which may hint at an alternative proof!