This one was going to be a cry for help and a request to confirm that I'm not going cuckoo, but I think I figured it out. Still, for giggles, hear this:

I was playing around with one of my favorite Himalayan barbershop (Galuga, one of those projects that just generate endless yak shaving) when I noticed something funny with tied filehandles. After much head-scratching, I was able to corner the problem with this small example:

#syntax: perl
use 5.10.0;

use strict;
use warnings;

{
    package Foo;

    sub TIEHANDLE { return bless \my $i, shift; }

    sub PRINT { shift; $::X .= join '', @_, $\; } 

}

say "this is perl $]\n";

print '>>>';
print "nothing here-->";
say   "but a CR there->";
print "and nothing here->";
print '<<<';

tie *::FOO, 'Foo';

print "\n\n";
print FOO "nothing here-->";
say   FOO "but a CR there->";
print FOO "and nothing here->";
print '>>>', $::X, '<<<';

$\ = '';
$::X = '';

print "\n\n";
print FOO "nothing here-->";
say   FOO "but a CR there->";
print FOO "and nothing here->";
print '>>>', $::X, '<<<';

Basically, I define a package that will act like a filehandle. The $\ used in its PRINT is, according to perltie, the magic way to make sure print and say will both work: in the case of a say, $\ will be localized and assigned a carriage return. Sounds good, no? But let's see what happens when we run this baby:

$ perl fit_to_be_tied.pl 
this is perl 5.014002

>>>nothing here-->but a CR there->
and nothing here-><<<

Use of uninitialized value $\ in join or string at tie.pl line 13.
>>>nothing here-->but a CR there->
and nothing here->
<<<

>>>nothing here-->but a CR there->
and nothing here-><<<

The first stanza using STDOUT looks okay, but for the second stanza the print after the say seems to cling to having $\ = "\n" even if it's not supposed to. And to add a dash of uh? to the whole thing, if we set the global '$' not to be undef (as we do in the third stanza), all is fine again.

Fortunately, and thanks to perlbrew, I have a perl 5.17.6 lying around, and if I try the same test with it, I get:

$ perl fit_to_be_tied.pl 
this is perl 5.017006

>>>nothing here-->but a CR there->
and nothing here-><<<

Use of uninitialized value $\ in join or string at tie.pl line 13.
Use of uninitialized value $\ in join or string at tie.pl line 13.
>>>nothing here-->but a CR there->
and nothing here-><<<

>>>nothing here-->but a CR there->
and nothing here-><<<

Well, look at that. Everything's now behaving the way it should. I quickly perused the perldeltas since 5.16.1 (where the issue can also be seen) but didn't noticed any bug fix that could account for the change of behavior. Well, even then I won't complain. I'll just have one more reason to way for 5.18.0 with bated breath...