product json updates

tests/test_brand_table_scores.py

@@ -223,9 +223,41 @@ def test_the_brand_match_is_case_insensitive(mirror):
     nutrition_score_sync.sync_scores_to_brand_tables()
 
     for sql, params in mirror.calls:
-        assert "lower(i.brand) = lower(%s)" in sql
+        assert "= lower(%s)" in sql
         assert params == ("Cadbury",)
         assert "i.brand = %s" not in sql
+        assert "brand = %s" not in sql
+
+
+def test_a_brand_spelled_two_ways_resolves_to_one_row(mirror):
+    """`nutrition_insights` really does hold both 'Grb' and 'GRB' for the same
+    eleven products, written by two enrichment runs a month apart and carrying
+    different scores. Folding the brand matches both, and `UPDATE ... FROM`
+    against a multi-row match picks an arbitrary one - so the mirror wrote a
+    different value on every run. Caught by re-running the sync and watching
+    rows change again, which is the whole reason the idempotence test earns
+    its place.
+    """
+    nutrition_score_sync.sync_scores_to_brand_tables()
+
+    copy_sql = mirror.calls[0][0]
+    assert "SELECT DISTINCT ON (image_id)" in copy_sql
+
+
+def test_a_real_score_beats_a_null_and_then_the_newest_wins(mirror):
+    """The tie-break, stated as an assertion.
+
+    A NULL is the absence of a measurement, not a measurement of absence, so
+    it must never overwrite a known score just by being more recent - the same
+    instinct as the COALESCE in the Excel upload's upsert.
+    """
+    nutrition_score_sync.sync_scores_to_brand_tables()
+
+    copy_sql = mirror.calls[0][0]
+    order = copy_sql[copy_sql.index("ORDER BY"):]
+    assert "(health_score IS NOT NULL) DESC" in order
+    assert "generated_at DESC" in order
+    assert order.index("health_score IS NOT NULL") < order.index("generated_at")
 
 
 def test_the_mirror_uses_is_distinct_from_so_a_second_run_is_a_no_op(mirror):