updates based on copilot

Author: tballison

tika-ml/tika-ml-junkdetect/src/main/java/org/apache/tika/ml/junkdetect/tools/TrainJunkModel.java

@@ -1161,7 +1161,8 @@ static float[] trainGlobalCombiner(JunkDetector fx,
 
                 // Mixed-encoding (korA-class) pair: coherent text with heavy
                 // undecodable FFFD widget bytes (~15-20% of high bytes) must BEAT
-                // coherent-looking mojibake of the same bytes.  As a contrastive
+                // coherent-looking mojibake of the same source text (the clean w,
+                // without the FFFD widgets).  As a contrastive
                 // PAIR (not a pointwise positive) it adds no class imbalance, and
                 // it forces z1/coherence to stay the discriminator: the correct
                 // side has MORE FFFD (z6 favors the wrong side) and equal z7

tika-ml/tika-ml-junkdetect/src/test/java/org/apache/tika/ml/junkdetect/JunkDetectorRoundTripTest.java

@@ -371,6 +371,9 @@ private static float expectedRunZ(BigramTables tables, String text, float mu, fl
                 acc[1] += 1;
             }
         });
+        if (acc[1] == 0) {
+            throw new IllegalArgumentException("no scorable bigrams for: " + text);
+        }
         return (float) ((acc[0] / acc[1] - mu) / sigma);
     }