Understanding eval

After many hours of debugging code, I came to the conclusion that I
have a fundamental misunderstanding regarding eval, and hope that
someone here can explain to me, why the following code acts as it does:

foo <- function(expr) {
  eval(substitute(expr), envir=list(a=5), enclos=parent.frame())
}
bar <- function(er) {
  foo(er)
}

 > foo(a)

[1] 5

 > bar(a)

Error in eval(expr, envir, enclos) : object "a" not found


Now, regarding the "bar(a)" call, this is my understanding of what
happens, hoping someone will correct me where I'm wrong.

1) bar is called. Its evaluation frame contains the association "er=a".
2) bar calls foo. So foo is called, and its evaluation frame contains
the association "expr=er", with enclosing environment the local
environment of bar.
3) foo calls eval.
4) eval starts by evaluating "substitute(expr)" in foo's environment.
"substitute" then locates expr in foo's environment, and replaces it
with er. So the result of this process is the symbol er, which is
what will now be evaluated by eval.
5) eval then creates the environment where this evaluation will take
place. It does that by creating an environment containing the frame
"a=5", and with enclosing environment the parent frame of foo, which
is bar's environment.
6) So, as I understand it, the symbol "er" is going to now be
evaluated in an environment where a is set to 5 and er is set to a,
along with whatever is in the user's workspace.
7) So the first step now is looking up a definition for er. Nothing
is found in the current frame, so the evaluation proceeds to bar's
environment, where the association "er=a" is found, so er is replaced
by a.
8) Now, and perhaps this is where I misunderstand things, the lookup
for a will take place. My thinking was that the lookup would start
from the evaluation environment that eval created, and hence would
locate the a=5 value. But this is clearly not what happens.

Anyway, hope someone will correct me where I'm wrong, and explain to
me what I am doing wrong, and ideally how to diagnose such things.

Thanks,
Haris Skiadas
Department of Mathematics and Computer Science
Hanover College