Adding context:
If $E_i\in\mathscr C$, let $F_i = e(E_i)\in\mathscr D$. Otherwise, $e$
was applied in reaching $E_i$, and so there exists $F_i \in \mathscr
D$ from which $E_i$ is reachable.
Of course, without the definitions of $E_i$, $e$, $\mathscr C$, and $\mathscr D$, this isn't much use either. We're given a configuration $C$. $e$ is by assumption an event applicable to $C$. $\mathscr C$ is the set of all configurations reachable from a given configuration $C$ where event $e$ has not been applied. $\mathscr D$ is all the configurations you get when you apply $e$ to the configurations in $\mathscr C$. $E_i$ is a(n $i$-valent) configuration reachable from $C$.
The first paragraph of the lemma shows that $e$ is applicable to all configurations in $\mathscr C$. Given this, if $E_i\notin \mathscr C$, then, by definition of $\mathscr C$ and the fact that $E_i$ is reachable from $C$, $E_i$ must be reachable via some configuration that applies $e$. Any run from $C$ to $E_i$ must have a first time where $e$ is applied. Any configuration before that is in $\mathscr C$ by definition. The configuration at the point $e$ is applied is in $\mathscr D$ by definition and the fact that $e$ is applicable to all configurations in $\mathscr C$. That configuration is $F_i$.
Again, your second quote lacks context. The "such steps" mentioned in it are the steps from a bivalent configuration to a univalent configuration. Avoiding "such steps" is only relevant if we must start from a bivalent initial configuration, which is what Lemma 2 establishes. Lemma 2 critically depends on allowing up to one faulty process. If the protocol did not have to work correctly in the face of up to one faulty process, then Lemma 2 would not apply.
The idea of Lemma 2 is that if we have two configurations that differ only in how the input register is set for a single process, then if they can both reach a decision without ever executing that process (i.e. that process is faulty), then they must reach the same decision, since, once we remove that process, they have the same set of processes with the same initial state receiving the same messages. If we assume there are no bivalent initial configurations (for contradiction), then by partial correctness it is necessarily the case that there are two configurations that would make different decisions, but differ from each other only by a single process' input. Thus a contradiction.
It's trivial to make a consensus protocol when no processes are faulty. Simply pick one process (via setting its initial state appropriately) to be the Decider. It copies its input to its output. All other processes do nothing. Clearly this scheme won't work if the Decider is faulty.
