Add rejected learning keywords to blocklist for QA before standup

Your learning keyword exports land with rejected entries, causing noisy training data and wasted reviewer time. Reviewers get a curated blocklist so QA closes faster before the next model retrain.