Is the connected sum of complex manifolds also complex?
EDIT: I am greatly appreciative of Aleksandar Milivojevic for pointing out the first paragraph of my original answer is false, and therefore does not give a proof of the following two paragraphs. (For instance, $K3 \# \overline{K3}$ supports no complex structure: Aleksandar kindly explained to me that this has signature zero but Euler characteristic $46$, and $\chi + \sigma$ is always divisible by $4$ on an almost complex 4-manifold.
I leave these here for posterity, separated from the correct content by lines. It remains an interesting question whether or not $M \# N$ can support a complex structure when $M$ and $N$ are singly-even dimensional complex manifolds. There is no obstruction coming from the nonexistence of an almost complex structure, so it is not totally impossible.
Let $M$ and $N$ be complex manifolds. Because we have a local form (there is a neighborhood of a point isomorphic to the the complex unit disc) you can verify that $M \# \overline N$ has a natural complex structure - just line up the copies of $\Bbb C^n \setminus \{0\}$ you're gluing together so the almost complex structures match up. Then the Nijenhuis tensor still vanishes.
So, if $N$ supports a complex structure, does $\overline N$? This would mean that $M \# N$ supports a complex structure. If $\text{dim}(N) = 4n+2$, then this is true: if $J$ is your (integrable almost) complex structure, then the complex conjugate $\overline J$ gives a complex structure on $\overline N$. The reason this doesn't work in dimension $4n$ is because $\overline J$ induces the same orientation as $J$!
So I wouldn't quite call it a fluke. It's true for every $4n+2$-dimensional complex manifold, not just $2$-dimensional ones. But it's not true in every dimension.
For 4-manifolds, it is a theorem of Wu that $M$ admits an almost complex structure with $c_1(J) = c \in H^2(M;\Bbb Z)$ if and only if $c$ reduces to $w_2$ mod 2, and $c^2 = 3\sigma + 2\chi$. You can prove using this criterion that a most two of $M, N$, and $M \# N$ can admit almost complex structures; in particular, $\Bbb{CP}^2 \# \Bbb{CP}^2$ cannot admit an almost complex structure.
I don't know about higher dimensions $4n$. It might be wise to think about almost complex structures instead of complex ones first, because those are easier to work with and there are no known high dimensional (complex dimension $> 2$) examples of almost complex manifolds that do not admit a complex structure.
$X = \Bbb{CP}^4 \#\Bbb{CP}^4$ does not support an almost complex structure. I suspect this is true for $\Bbb{CP}^{2n} \# \Bbb{CP}^{2n}$ for any $n>0$; I can't prove it. The general idea is to obtain a "higher-dimensional" Wu theorem on which manifolds admit complex structures in terms of various characteristic classes of your manifold. I think this is possible in all dimensions, but probably becomes absurd very quickly. The 8-dimensional version is here. This is what we will use.
Preliminaries: the cohomology ring of $X$ is the direct sum of two copies of the cohomology ring of $\Bbb{CP}^4$, identifying the unit and volume form in each copy. Denote the two generators of $H^2(X;\Bbb Z)$ as $x_1, x_2$.
1) Wu classes. Because we know precisely what the cohomology ring is, it's easy to calculate this. In particular, $Sq^2(x_1^3) = x_1^4$, and $Sq^2(x_2^3) = x_2^4$, so $v_2 = x_1+x_2$. Similarly we see that $Sq^4(x_1^2) = x_1^4, Sq^4(x_2^2)=x_2^4$, so $v_4 = x_1^2+x_2^2$. Trivially $v_6 = 0$.
2) Stiefel-Whitney classes. Because the total SW-class $w$ satisfies $Sq(v)$, where $Sq$ is the total Steenrod square and $v$ is the total Wu class, we obtain $w_2 = v_2 = x_1+x_2$, $w_4 = v_2^2+v_4 = 0$, $w_6 = Sq^2(v_4) = 0$.
3) Pontryagin classes. We can give $X$ a cell structure such that its 4-skeleton is $\Bbb{CP}^2 \vee \Bbb{CP}^2$. Because $X_4 \hookrightarrow X$ induces an isomorphism on $H^4$, the Pontryagin class is determined by the restriction of the classifying map $X \to BSO(8)$ to $X_4$. We can choose this classifying map to be of the form $f \vee f$, where $f$ is the restriction of the classifying map of $\Bbb{CP}^4$ to its 4-skeleton. Because we know $p_1(\Bbb{CP}^4) = 5x_1^2$, we see that $$p_1(X) = 5x_1^2+5x_2^2.$$ Now the Hirzebruch signature theorem tells us that $p_2 = 20x_1^4$.
Now let's plug into the theorem. Our cohomology classes $u, v$ would have to be of the form $u = (2k+1)x_1 + (2\ell+1)x_2$, and $v = 2mx_1^3 + 2nx_2^3$, where $m,n$ have the same parity (by b, c).
Then because $\chi(M) = 16$, d) takes the form $$128 = 80 + 32(km+\ell n) + 16(m+n) - (2k+1)^4 - (2\ell+1)^4 + 10(2k+1)^2 + 10(2\ell+1)^2 - 50.$$
Simplifying we get $$80 = 32(km + \ell n - k^3 - \ell^3 + k^2 + \ell^2) + 16(k+\ell - k^4 - \ell^4 +m +n)$$ Reducing mod 32 and remembering that $m+n$ is even we obtain $$16 \equiv 16(k+\ell - k^4- \ell^4) \mod 32$$ But $k+\ell$ is odd iff $k^4 + \ell^4$ is odd, so the right side is $0 \mod 32$. This is a contradiction, as desired.
That this was so much work suggests that this is, uh, the wrong approach to prove that $\Bbb{CP}^{2n} \# \Bbb{CP}^{2n}$ never supports a complex structure. Maybe someone more gifted with obstruction theory than I can prove this.