You did better than I did (_especially_ with optimizing sub a, and with getting rid of the Socket constants, which I knew was possible but didn't do), although you did miss a couple of opportunities...

• As someone else also noted, undef$/, while it saves one stroke over the original, is still suboptimal. I have $/=$3. A related situation occurs later with undef$k{$p}, only there $3 might be defined, so I have $k{$p}=$7 in that case. $7 will always be undef in molester. Update: your updated version does this in the first of the two cases.
• open F,'>',$_[1] is slightly suboptimal. I have open F,">$_[1]", saving one stroke. I saved another stroke doing the same thing when it's opened for reading. Update: your updated version does this too.
• Because sub a is declared before sub t, you can dispense with the parens when t calls a, saving two characters: (connect X,a$w) Update: your updated version does even better than this.
• My /. post doesn't _demonstrate_ this, but it mentions it: converting the if/else isn't as fruitful as I'd hoped, due to the need to use and after the print (for precedence reasons), but it does shave off a couple of strokes. Update: I believe this is still good for a couple of strokes.
• another update: I think we can shave one more stroke by using a statement modifier in sub g. This will cause F to be closed whether it's successfully opened or not, but that should be okay I _think_.
• Yet Another Update: But I lost that character again fixing what I think is a bug in your regex optimization, having to do with the slashes that separate commands.

We also both missed the argumentless-close trick that someone else pointed out, although I would have to test that one to be sure it doesn't close the socket also (or that doing so doesn't change the operation of the script). Can someone confirm which close statements need their arguments and, for that matter, whether it's possible to eliminate any of the close statements altogether without breaking something?

So I threw out my version and worked from yours update: again, and this is what I now have, weighing in at 508 bytes not counting line breaks (522 bytes with a full ASCII CRLF ending every line but the last, or 515 with Unix-style one-character newlines):

$/=$_;$(=shift;$a=shift;i(shift);socket S,2,1,6;bind S,a($a);listen
S,5;while(@ARGV&&($_="$( $a f".shift)||accept(C,S)&&($_=<C>)&&close
C){m!^(.*?) (.*?) ([e-i])([^/]*)./!s&&$(eq$1&&&$3($2,$4,$')}sub f{t($_
+,@_)for
keys%k}sub a{pop=~/:/;pack'CxnC4x8',2,$',split'\.',$`}sub i{$k{$_[0]}=
+1}sub
g{t($_[0],$a,"e$_[1]",<F>)if open F,"<$_[1]";close F}sub h{t($_[0],$_,
+'i')for
keys%k}sub t{socket X,2,1,6;(connect X,a($w=shift))?print
X"$( $_[0] $_[1]/$_[2]"and close X:$k{$w}=$7}sub
e{open F,">$_[1]";print F $_[2];close F}
[download]

Note that this code is completely, utterly untested.