Pick a t.judge threshold and correct a reported rate

Both parts reuse the previous exercise's gold set (the same 80 generated items, your labels and the same splits), so every number here is measured against your own judgement. Part A treats t.judge's score as what the docs say it is, an uncalibrated estimate, and lets your labels decide where "pass" starts. Part B starts from the frozen v3 judge's test-split TPR and TNR and asks what pass rate a solicitor would have seen. The decision that matters most is in Part B's bootstrap: resample the labelled test rows, not just the fresh batch, because the uncertainty in TPR and TNR is usually the bigger part of the interval.