I am reading the "Reaching Agreement in the Presence of Faults", M. Pease et al and trying to understand their proof for the $n \geq 3m+1$ case.
In the induction step $m \gt 0$ it says the following:
... Any other set satisfying these requirements, moreover, must contain a nonfaulty processor (since it must be of size $\gt \frac{(n + m)}{2}$, and $n \ge 3m + 1$) and must therefore also yield $V_q$ as the common value. The algorithm thus terminates at step (1), and $p$ records $V_q$ and $q$ as required...
This is not obvious to me and seems wrong. For example, assume $Q$ is a set of processors of length $\ \ge \frac{n+m}{2}$ where only one is a faulty processor $f$ and rest are nonfaulty. Now we consider all strings $w$ of length $\le m$ of this set and in particular $\sigma_p(pfq)$. Since $f$ is a faulty processor it can return any value, not necessarily $V_q$, thus there cannot be a common value between all $w$ strings. This means that any set $Q$ that contains at least one nonfaulty processor cannot be used at all.
Is my understanding wrong?