# Closure of a CFL under specific operation

Consider the following operation on language $L$:

$\mathrm{inv}(L) = \{ xy^Rz \mid x,y,z\in \Sigma^*, xyz\in L \}$

I understand that if $L$ is regular, then $\mathrm{inv}(L)$ is regular too, and proved it by guesing when $y^R$ starts and running it on the inverse DFA. However, if $L$ is a CFL, then $\mathrm{inv}(L)$ is not, and I don't understand why. Can't we just also guess when $y^R$ starts, insert all of it into the stack, then simulate the DFA of $L$ on each item we take out, then continue on $z$ when it's empty?

thanks.

You are right, regular languages are closed under inversion $\mathrm{inv}$, using some proper guesses, or see my answer to a relevant question here.
Context-free languages are not closed under $\mathrm{inv}$. Your intuition does not work. We cannot push $y$ on the stack because it will make the 'real' stack below it inaccessible for the simulation.
Here is a possible counter example. Consider $L = \{ a^nb^nc^md^m \mid m,n\ge 1 \}$. Then try $\mathrm{inv}(L) \cap a^*c^*b^*d^*$.