Now that you've drawn my attention to the code again. I say it is that baroque. My initial justification was just for the  and 1 stuff which I still don't find so bad.

I would still like a or $a cmp $b or a or 0 or  die "domain err". It is not readily apparent what is to happen with other/other comparisions. This seems like a significant decision point in the code (by the identifiers) and so a good place for extra clarity. I can imagine myself referring to this snippet to understand how all the subsequent parsing code is recieving data. This may be such a recurring problem in the codebase that it seems unimportant.

What does irk is the inconsistent use of short-circuiting, the sub-clauses are unnecessary and so confusing. (I just trusted that aspect the first time around.

If I were going to write in that form I'd write something like:

@List = sort {
    $a eq "one" and -1
 or $b eq "one" and 1
 or $a eq "two" and -1
 or $b eq "two" and 1   # no need to check if $a eq one
 or $a eq "thr" and -1
 or $b eq "thr" and 1
 or 0                   # advertising the default case
}
[download]

code block given to sort must behave (as POD stays.)
Namely, once f(a,b) return false f(b,a) must return true.
Some versions of perl dumped core when fed some misbehaving comparision sub.
Such core dumps were fixed, but general rule remain: comparision sub must behave

This means that sort {rand(5) <=> rand(5)} is obviously wrong, but sort {1} also not good.

The comparison function for sort must not just return 'true' and 'false'.
It must return '-1', '0' and '+1' for comparing 'Less Than', 'Equal To' and 'Greater Than', just like the two standard comparison operators 'cmp' and '<=>' (for stringwise and numeric comparison respectively).

Just using 'true' (ie non-zero, probably 1) and false (ie 0 or '') will lead to confusion.

If f(a, b) is 'false' then f(b, a) must also be false because a 'Equals' b - for whatever value of 'Equals' this sorting choses to use.
If f(a, b) is 'true' then f(b, a) must return '-1' (ie a true value, but *not* just true) because a is 'Greater Than' b in this case, and therefore b must be 'Less Than' a - not 'Equal To', or something is decidedly wierd.

You have to be consistent however you chose to do it.

You're right, thanks for correction.

I just could not resist to point out for general idea (I saw that in p5p list) but I forgot to recheck all details

But once we decided to be precise, let me make additional correction.
Requirement to return '-1', '0' and '+1' is too restrictive. According to perldoc -f sort:

parison order.  If SUBNAME is specified, it gives the name of a
subroutine that returns an integer less than, equal to, or
greater than 0, depending on how the elements of the list are
to be ordered.  (The "<=>" and "cmp" operators are extremely
useful in such routines.)  SUBNAME may be a scalar variable
[download]

So any negative and positive values are ok.

Best regards,
Courage, the Cowardly Dog

To clarify, prior to perl 5.005 (!) sort used to use libc's qsort, which would sometimes coredump on certain platforms with pathological return values. But this isn't a concern any longer unless you are stuck with a really old perl.