Let $M\models T$, a stable theory. Let $a,b\in U\succ M$. Are the following equivalent?
$\varphi(M,b)\neq\varnothing$ for every $\varphi(x,y)\in \color{red}{L(M)}$ such that $\varphi(a,b)$
$\varphi(M,b)\neq\varnothing$ for every $\varphi(x,y)\in \color{red}{L}$ such that $\varphi(a,b)$
No, parameters from $M$ can be important for witnessing dependence over $M$. Here's an example.
Let $M$ be an infinite $k$-vector space and $U$ a proper elementary extension. Let $v\in M$ with $v\neq 0$, let $b\in U\setminus M$, and let $a = b+v$. Let's observe two facts:
Let $\varphi(x,y)$ be the $L(M)$ formula $x = y + v$. Then $\varphi(a,b)\in \text{tp}(a/Mb)$, but is not satisfiable in $M$. Indeed, $a$ is its only realization in $U$, and $a\notin M$.
On the other hand, every $L$-formula in $\text{tp}(a/b)$ is satisfiable in $M$. In fact, every non-zero element of $M$ realizes the whole type! Indeed, for any non-zero $a'\in M$, $a'$ and $b$ are linearly independent (since $b\notin M$), so there is an automorphism of $U$ fixing $b$ and moving $a$ to $a'$.