Suppose we have a linear IV model with: $$ y_i=\beta x_i + \epsilon_i$$ $$x_i=\gamma z_i + u_i $$with $E(x_i \epsilon_i)\neq 0$ and $E(z_i \epsilon_i)=0$.
Then we can estimate the residuals in the reduced form $\hat{u}_i$ and run the regression: $$ y_i=\beta x_i + \alpha \hat{u}_i + v_i$$
which yields a consistent estimator for $\beta$ (equivalent to 2SLS).
We can run the test for $H_0: \alpha = 0$ using a standard t-test.
Would this be a valid test to check the validity of the instrument (i.e. $H_0: E(z_i \epsilon_i)=0$)? I've read that it's not possible to test for instrument validity but I don't understand why this test is not acceptable. What is this testing for?