How do you figure that out?

First, don't put named subs inside loops or other subs. It makes no sense. The latter will even warn ("will not stay shared") if it captures anything from the sub its in.

Then the answer is simple: All variables declared before it in a scope that hasn't been closed.

my $x;                <- this one.

{
   my $y;             <- not this one.
   ...
}

{
    my $z;            <- this one.
    sub f { ... }
}
[download]

For anonymous subs, it's a different story. Feel free to use them in loops and other subs. They'll pick up the variables that are visible when the sub is executed.

And why does a my variable that is in scope inside the loop's block

for localizes its iterator variable. Localizing is temporary replacing a variable with a new one.

You captured $num before the loop started. This is different from the $num of the first pass, which is different from $num of the second pass, ..., which is different from $num of the fifth pass.

(Ok, I lied. I believe an optimisation will actually cause the variable of the first pass to be reused for the subsequent passes, but that's transparent here.)

my $x;                <- this one.

{
   my $y;             <- not this one.
   ...
}

{
    my $z;            <- this one.
    sub f { ... }
}
[download]

A minor quibble: We can't entirely ignore intervening code. For example,

my $x = 'shadowed';      # <- invisible
{
    my $x = 'shadowing'; # <- visible
    sub f { say "I only see the \$x that's $x." };
}
[download]