Constructing a POVM to discriminate $m$ quantum states. What if they're linearly dependent?

A crucial hypothesis is missed in your construction.

Each $\phi_i$ must also satisfy $\phi_i \not \perp \psi_i$, otherwise $\langle \psi_i |E_i \psi_i\rangle >0$ is false.

This point provides an answer to your last question as well.

If $\psi$ is an added further vector, linearly dependent on the vectors $\psi_i$, the construction you made cannot be re-proposed as the constraint I pointed out cannot be satisfied. Indeed, even if the correspondingly added normalized vector $\phi$ is orthogonal to all $\psi_i$, it is (evidently) impossible that $\phi \not \perp \psi = \sum_{i=1}^n c_i \psi_i$.