Reverend,

Can you tell us the method you used to get the memory figures ? And your platform/perl details too ?
I ran your snippets on my RH9 with perl, v5.8.0 built for i386-linux-thread-multi and I made one change to snippet C to

 perl -le '%hash = (1 .. 1000000); for (keys %hash) { print ">"; <STDI
+N> ; exit}'
[download]

to ease the snippets death.

I just used top to look at the Size column, and my results are

COUNT	a mem	b mem	c mem	a-b change	a-c change
100,000	10,960	11,984	11,004	1,024 (9.3%)	44 (0.4%)
1,000,000	94,324	101,000	94,460	6,676 (7.1%)	136 (0.1%)

Note the top in RH9 truncates snippet B, 1,000,000 items to 101M, and I haven't figured out how to change the units for this field - perhaps I need to use something other than top...

So in conclusion, I agree, it does appear there is some optimisation going on - is this an example of lazy evaluation ?