Print the reference above the prediction

The old line gave a WER and the first 50 characters of the hypothesis, so
you could see how wrong an utterance was but not how. Printing the two in
full on consecutive lines shows the failure mode, which on this pipeline is
the interesting part: the prediction tracks the reference's rhythm and
phrase shape while the content drifts, because a prosody contour carries no
phonetic evidence for which words filled it.

Both lines are full length now -- the truncation was hiding the ends, which
is where the two most often diverge.

evaluate_dataset.py

@@ -327,11 +327,18 @@ def main(argv: list[str] | None = None) -> int:
                     hypothesis_transform=transform,
                 )
                 stats = word_errors(reference, result.text)
+                # Reference above, hypothesis directly below it, aligned.
+                # A WER on its own says how wrong; only the two lines
+                # together say in what way -- and on this pipeline that is
+                # the interesting part, since the words are usually the
+                # right shape and the wrong content.
                 print(f"[{index:>5}] WER {100 * stats['wer']:5.1f}%  "
-                      f"running {running:5.1f}%  | {result.text[:50]}",
+                      f"running {running:5.1f}%\n"
+                      f"        ref  {reference}\n"
+                      f"        ours {result.text}",
                       flush=True)
             else:
-                print(f"[{index:>5}] {result.text[:60]}", flush=True)
+                print(f"[{index:>5}] {result.text}", flush=True)
 
             if handle is not None:
                 # Same field names and order as the published runs'