Binary classifier over (text, rule-in-natural-Arabic).
seen wording 0.9929 acc / 0.9923 macro-F1
unseen wording 0.9799 acc / 0.9783 macro-F1 (majority baseline 0.6401)
minimal pairs 1080 @ 100%, flip-rate 1.00
Includes the full pipeline: paraphrase generation, data prep, training,
counterfactual eval, and the rule checkers that define the labels.
