Judge calibration

• Live: false
• n: 20
• Accuracy: 85.0% (17 / 20)
• Precision (ACCEPT=positive): 83.3% (10 / 12)
• Recall: 90.9% (10 / 11)
• F1: 0.870
• Confusion TA/FA/FR/TR: 10/2/1/7
• Caveat: n=20 (skipped PENDING=0) is too small for a published judge-accuracy claim; source=canned judgeOutput; does not affect production eval or the quality gate