use bignum;

$foo = 12345.6789;
$bar = $foo->copy()->bfround(0)->bfround(-5);
$baz = $foo - $bar;
print "what's wrong here? $foo - $bar = $baz\n";

$byhand = 12345.6789 - 12345;
print "expected $byhand\n";
[download]

Gives:

what's wrong here? 12345.6789 - 12346.00000 = -0.32110
expected 0.6789
[download]

Seems to me that the precision is propagated to the result value, and -0.32 rounds up to zero in your original example.