fix(m3-security-fpr-metric): recompute fpr_safe_items as combined FPR across all safe (answer=No) items

fpr_safe_items was 0.125 (near-miss-only FPR). Recomputing from actual scored
calibration data over all safe items (near-misses + off-target distractors)
gives 0.453 = 34 FPs / 75 trials. Per-class FPR rates retained as separate
manifest fields (fpr_near_miss=0.125, fpr_off_target_distractor=0.829).
Manifest item_counts still exactly match decisive JSONL (70 items). No item
code or labels changed.