The tests were run on a PowerMac G5 running Linux, PowerMacs are big-endian.

I reduced the issue to this program where I get the incorrect answer 25132:

{-# LANGUAGE BangPatterns #-}
{-# LANGUAGE MagicHash #-}
{-# LANGUAGE UnboxedTuples #-}

module Main where

import GHC.Exts

main :: IO ()
main = do
    putStrLn $ show (apply2 plusInt16# (50) (50))

apply2 :: (Int16# -> Int16# -> Int16#) -> Int -> Int -> Int
apply2 op (I# a) (I# b) =
    let (# sa, sb #) = (# narrowInt16# a, narrowInt16# b #)
        r = op sa sb
    in I# (extendInt16# r)
{-# NOINLINE apply2 #-}

Curiously, when I inline plusInt16# and remove the op parameter the result is correct.

The test passes on little-endian PowerPC 64-bit.

Note: This bug is different from #16222 (closed), which deals with the C calling convention.