I mean, dgrozev specifically points out that he proves $|P(t)-2019| \le 1$ and that an extension needs to be made to $|P(t)-2019| < 1$. He said he can provide that extension, but didn't. So I don't see how you can call his methods "wrong"; he just didn't provide the complete proof.
In any event, I have a proof that shows that at least $50\%$ of all large $d$ have such a polynomial. Clearly, from dgrozev's proof, there is only an issue if there are some $\epsilon_0,\epsilon_1,\dots,\epsilon_d \in \{\pm 1\}$ with $\epsilon_dt^d+\dots+\epsilon_1 t+\epsilon_0 = 2018$. Now, I claim that if there are such $\epsilon_0,\dots,\epsilon_d$, then there won't be $\epsilon_0',\dots,\epsilon_{d+1}'$ with $\epsilon_{d+1}'t^{d+1}+\dots+\epsilon_1't+\epsilon_0' = 2018$.
To show this claim, it suffices to show that two polynomials with only $\pm 1$ coefficients that are one degree apart have no common roots. It of course suffices to show that any two such polynomials have gcd $1$ over $\mathbb{Q}[X]$, which is equivalent to showing any two such polynomials share no common factor in $\mathbb{Q}[X]$, which essentially by Gauss's lemma (or its proof) is equivalent to two such polynomials sharing no common factor in $\mathbb{Z}[X]$. So suppose for the sake of contradiction that they did share a common factor in $\mathbb{Z}[X]$. This common factor must, of course, be monic or have first coefficient $-1$. Then, reducing everything mod $2$, $x^{d+1}+\dots+x+1$ and $x^d+\dots+x+1$ would share a nontrivial common factor in $\mathbb{F}_2[X]$, but they don't since $x^{d+1}+\dots+x+1-x(x^d+\dots+x+1) = 1$.