Or perhaps the "create a new instance" actually happens when the scope is left rather than when it is re-entered.

That's what a monk told me. He even told me the name of the macro, but it escapes me. Unfortunately, I lost the notes I took.

( Update: I did some digging in the Perl sources. Opcode pp_padsv (called to fetch an SV from the pad) adds a SAVEt_CLEARSV "directive" on the stack (if I understand correctly). When the scope is exited, leave_scope is called. When leave_scope encounters the SAVEt_CLEARSV "directive", it either reinitializes the SV (using SvOK_off) or it creates a new SV to replace the still-referenced one on the pad. Similar events occur for AVs and HVs. )

The following snippet demonstrates it's the refcount at end of the loop that matters, thus demonstrating the initialization happens at the end of the loop:

my @a;

for my $i (1..8) {
   foreach ('once') {
      my $x;
      push @a, \$x if $i >= 4;
      print(\$x, "\n");
   }

   pop @a if $i >= 7;
}
[download]

Output:

SCALAR(0x226024)
SCALAR(0x226024)
SCALAR(0x226024)
SCALAR(0x226024)
SCALAR(0x2252e0)
SCALAR(0x2252f8)
SCALAR(0x225334)  <-- different
SCALAR(0x22513c)  <-- different
[download]

If it was the ref count on entering the loop that mattered, the output would be:

SCALAR(0x226024)
SCALAR(0x226024)
SCALAR(0x226024)
SCALAR(0x226024)
SCALAR(0x2252e0)
SCALAR(0x2252f8)
SCALAR(0x2252f8)  <-- same as prev
SCALAR(0x2252f8)  <-- same as prev
[download]

Unfortunately, PadWalker can't examine the loop's pad from outside of the loop (although it can examine a sub's pad from outside the sub).