feat(vocab): implement SRS pair fingerprinting and deduplicate course-wide SRS items

backend/migrations-active/20260828000000-dedupe-course-wide-srs-items.cjs

@@ -0,0 +1,50 @@
+'use strict';
+
+/**
+ * Legacy SRS keys included lesson_id while the course pool used NULL. The same
+ * word pair could therefore receive parallel schedules. Keep the most advanced
+ * record, preserve its earliest due time, and remove only exact course-wide
+ * duplicates.
+ */
+module.exports = {
+  async up(queryInterface) {
+    await queryInterface.sequelize.query(`
+      WITH grouped AS (
+        SELECT
+          id,
+          FIRST_VALUE(id) OVER bucket AS keeper_id,
+          MIN(next_due_at) OVER bucket AS earliest_due_at,
+          MAX(correct_count) OVER bucket AS max_correct_count,
+          MAX(wrong_count) OVER bucket AS max_wrong_count,
+          MAX(lapse_count) OVER bucket AS max_lapse_count,
+          BOOL_OR(is_hard) OVER bucket AS any_hard,
+          ROW_NUMBER() OVER bucket AS row_number
+        FROM community.vocab_srs_item
+        WINDOW bucket AS (
+          PARTITION BY user_id, course_id, UPPER(direction),
+            LOWER(REGEXP_REPLACE(learning, '[[:punct:][:space:]]+', ' ', 'g')),
+            LOWER(REGEXP_REPLACE(reference, '[[:punct:][:space:]]+', ' ', 'g'))
+          ORDER BY correct_count DESC, last_reviewed_at DESC NULLS LAST, id ASC
+          ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING
+        )
+      ), updated AS (
+        UPDATE community.vocab_srs_item AS item
+        SET next_due_at = grouped.earliest_due_at,
+            correct_count = grouped.max_correct_count,
+            wrong_count = grouped.max_wrong_count,
+            lapse_count = grouped.max_lapse_count,
+            is_hard = grouped.any_hard
+        FROM grouped
+        WHERE item.id = grouped.keeper_id
+        RETURNING item.id
+      )
+      DELETE FROM community.vocab_srs_item AS item
+      USING grouped
+      WHERE item.id = grouped.id
+        AND grouped.row_number > 1;
+    `);
+  },
+  async down() {
+    // Duplicate rows are intentionally not recreated.
+  }
+};