fix: WER=0 bug when no segments match in evaluation

When zero predicted segments overlap with gold segments in time,
summarise_matches reported WER=0.0 instead of WER=1.0 (due to
0/max(0,1)=0 division). This made grid search pick "no overlap"
configs as winners, corrupting all subsequent parameter iterations.

Fix: WER=1.0 when no matches. Also add recall-weighted WER/CER
that penalizes missed gold segments (if only 50% of gold matched,
even with perfect matched WER, overall WER = 0.5*0 + 0.5*1 = 50%).
Add matched_wer, matched_cer, recall to summary output.

Made-with: Cursor

evaluation.py

@@ -376,18 +376,36 @@ def summarise_matches(
     only_pred = [m for m in matches if m.get("gold") is None]
     only_gold = [m for m in matches if m.get("pred") is None]
 
-    avg_wer = sum(m["wer"] for m in matched) / max(len(matched), 1)
-    avg_cer = sum(m["cer"] for m in matched) / max(len(matched), 1)
-    speaker_correct = sum(1 for m in matched if m.get("speaker_match")) / max(len(matched), 1)
+    total_gold = len(only_gold) + len(matched)
+
+    if matched:
+        matched_wer = sum(m["wer"] for m in matched) / len(matched)
+        matched_cer = sum(m["cer"] for m in matched) / len(matched)
+        speaker_correct = sum(1 for m in matched if m.get("speaker_match")) / len(matched)
+    else:
+        matched_wer = 1.0
+        matched_cer = 1.0
+        speaker_correct = 0.0
+
+    if total_gold > 0:
+        recall = len(matched) / total_gold
+        avg_wer = matched_wer * recall + 1.0 * (1.0 - recall)
+        avg_cer = matched_cer * recall + 1.0 * (1.0 - recall)
+    else:
+        avg_wer = matched_wer
+        avg_cer = matched_cer
 
     summary = {
         "total_predicted_segments": len(only_pred) + len(matched),
-        "total_gold_segments": len(only_gold) + len(matched),
+        "total_gold_segments": total_gold,
         "matched_segments": len(matched),
         "unmatched_pred": len(only_pred),
         "missed_gold": len(only_gold),
         "avg_wer": round(avg_wer, 4),
         "avg_cer": round(avg_cer, 4),
+        "matched_wer": round(matched_wer, 4),
+        "matched_cer": round(matched_cer, 4),
+        "recall": round(recall, 4) if total_gold > 0 else 0.0,
         "speaker_accuracy": round(speaker_correct, 4),
     }