Fix the threshold picker for a rare defect model

An AI trained a defect detector on data with 1 defect in 200, using an unweighted loss and a fixed 0.5 threshold. The loss fell smoothly, and the model almost always predicted "no defect". Part of the fix is to choose the decision threshold on validation data by F1, a score that balances precision and recall and ignores the easy true negatives.

An assistant wrote select_f1_threshold(scores, labels, thresholds). It is in the editor. It has a bug. Find it and fix it.

Input and output:

• scores: the model's score for each validation row.
• labels: each row's label, at the same index: 1 for a defect, anything else for none. The two lists have the same length.
• thresholds: the thresholds to try, in order.
• Return the best threshold, or -1 if there are no candidates.

Rules the tests check:

• At a threshold, a row is predicted positive when its score is at least the threshold.
• Count the true positives (predicted positive, label 1), the false positives (predicted positive, label not 1) and the false negatives (not predicted positive, label 1).
• F1 is 2 times the true positives, divided by 2 times the true positives plus the false positives plus the false negatives. If that denominator is 0, F1 is 0.
• Return the threshold with the highest F1. On a tie, return the one that comes first in thresholds.