Two wrongs apparently do make a right. First, you’re using bad coordinates for the event horizon, and second, as a result, you divided by $0$ by plugging in $r=2m$. But, somehow you got that the induced metric tensor is $g=-r^2\,g_{S^2}$, which actually is the correct induced metric on the event horizon of Schwarzschild. But unfortunately, you made a third error and concluded that this is spacelike, when in fact this says the induced metric is degenerate, and thus the hypersurface is null (keep in mind a hypersurface is 3-dimensional in this case, so a non-degenerate metric would at the very minimum see three coordinates, but you only see two, so it’s definitely degenerate).
Correct Way to Analyze the Spacetime.
As mentioned already, you should use a good set of coordinates which cover the event horizon. One example is by the Eddington-Finkelstein coordinates $(v:=t+r^*,r,\theta,\phi)$, in which the metric is (with my signature, which is the opposite of yours, because I don’t want to type something wrong later on out of habit)
\begin{align}
g&=-\left(1-\frac{2m}{r}\right)\,dv^2+2\,dv\,dr+r^2g_{\Bbb{S}^2}.
\end{align}
This is valid for $(v,r)\in \Bbb{R}\times (0,\infty)$ and $(\theta,\phi)$ any spherical coordinates on $\Bbb{S}^2$. In particular, this covers the Schwarzschild exterior and also the black hole region. Hence, now we can set $r=2m$ (i.e consider the pullback of this metric to the hypersurface $\mathcal{H}_+:=\{(v,r,\theta,\phi)\,:\, r=2m\}$) to get
\begin{align}
g_{\mathcal{H}_+}&=0+0+4m^2\,g_{\Bbb{S}^2}=4m^2g_{\Bbb{S}^2}.
\end{align}
This is a degenerate metric on the 3-dimensional manifold $\mathcal{H}_+$, thereby proving that the event horizon is a null hypersurface.
By the way, another fun feature to notice is that the coordinate vector field $T:=\left(\frac{\partial}{\partial v}\right)_{(v,r,\theta,\phi)}$ in the $(v,r,\theta,\phi)$ coordinate system agrees with the vector field $\left(\frac{\partial}{\partial t}\right)_{(t,r,\theta,\phi)}$ from our original $(t,r,\theta,\phi)$ coordinate system (see here for the detailed calculation); in fact it is the unique analytic extension (but analyticity isn’t really the way to judge things; the key is that this $T$ is Killing, and timelike for large $r$, in fact for $2m<r<\infty$). Now, we can observe that $g(T,T)=g_{vv}=-\left(1-\frac{2m}{r}\right)$, so with my sign convention, $T$ is spacelike, null, timelike according as to whether $r\in (0,2m), r=2m, r\in (2m,\infty)$ respectively. So, our usual notion of time is getting more and more null as we approach the event horizon (geometrically, the timecones are “tilting”, until their boundary, which is the null cone, becomes tangent to the event horizon).
Equivalence of Definitions of Null Hypersurface
There are actually (atleast two) ways to describe a null hypersurface:
Lemma.
Let $(V,g)$ be a Lorentzian inner product space and $H\subset V$ a codimension-1 subspace. Then, $H$ is degenerate if and only if the orthogonal complement $H^{\perp}$ is spanned by a non-zero null vector (i.e the normal is null).
In either of these cases, we say $H$ is a null hyperplane.
To prove the $(\implies)$ direction, the meaning of degeneracy is that there exists a non-zero vector $\zeta\in H$ such that $g(\zeta,\cdot)|_H=0$. In particular, this says $\zeta$ is orthogonal to $H$, and since $\zeta\in H$, it says $g(\zeta,\zeta)=0$, and so $\zeta$ is null, i.e $H$ has a null normal.
Conversely, suppose $\zeta$ is a (non-zero) null normal. By virtue of being null, we have $g(\zeta,\zeta)=0$, meaning $\zeta\in (H^{\perp})^{\perp}=H$ (the inclusion $H\subset H^{\perp\perp}$ is true for any subspace by simple definition unwinding; the reverse inclusion follows from the rank-nullity theorem which implies these two subspaces have the same dimension, and thus must be equal). Thus, we have a non-zero vector $\zeta\in H$ which is orthogonal to $H$, i.e $g(\zeta,\cdot)|_H=0$, thus proving that $H$ is a degenerate subspace relative to $g$.
If you apply this argument at each tangent space, then you of course get the corresponding statement at the level of Lorentzian manifolds and null hypersurfaces. So, above, we saw directly that $g_{\mathcal{H}_+}$ was a degenerate metric and concluded from here that the hypersurface is null. Alternatively, we can see that the vector field $T=\frac{\partial}{\partial v}$ is such that
- it is obviously nowhere-vanishing (it’s a coordinate vector field afterall)
- $g_{vv}=0$ (i.e it is null). Said again, $T$ is null and orthogonal to $\frac{\partial}{\partial v}$.
- $g_{v,\theta}=g_{v,\phi}=0$, so $T$ is orthogonal to $\frac{\partial}{\partial\theta}, \frac{\partial}{\partial\phi} $.
Since $\frac{\partial}{\partial v}, \frac{\partial}{\partial\theta} , \frac{\partial}{\partial\phi} $ are tangent to the hypersurface $\mathcal{H}_+=\{r=2m\}$, and span the whole tangent spaces of the hypersurface, it follows that $T$ is orthogonal to the hypersurface. Thus, we see according to our second definition that $\mathcal{H}_+$ is a null hypersurface.
A third way to see/calculate that the hypersurface is null is to again use the Eddington-Finkelstein coordinates as above, and explicitly compute the normal vector, $\text{grad}_g(r)$ given by the gradient of the function $r$, or simply the inner product with itself (since we know that is the normal, and we only want to know its causal nature):
\begin{align}
g(\text{grad}(r),\text{grad}(r))&=g^{ab}\frac{\partial r}{\partial x^a}\frac{\partial r}{\partial x^b}=g^{rr},
\end{align}
since we’re dealing with coordinate vector-fields. So, we just need to invert the suitable matrix and extract this $rr$ component; a short calculation reveals it is $1-\frac{2m}{r}$, so when $r=2m$, it is a null vector field. Hence, the normal to $\mathcal{H}_+$ is null, so once again proving the hypersurface is null.