Update smoketest to account for backfills (#1882)

As I wrote in the commend:
We backfill island areas with data from the 2010 census, so if THOSE tracts
have data beyond the data source, that's to be expected and is fine to pass.
If some other state or territory does though, this should fail

This ends up being a nice way of documenting that behavior i guess!

data/data-pipeline/data_pipeline/tests/score/test_output.py

@@ -7,6 +7,7 @@
 import numpy as np
 from data_pipeline.score import field_names
 from data_pipeline.score.field_names import GEOID_TRACT_FIELD
+from data_pipeline.etl.score.constants import TILES_ISLAND_AREA_FIPS_CODES
 from .fixtures import (
     final_score_df,
     ejscreen_df,
@@ -287,7 +288,19 @@ def test_data_sources(
 
         # Make sure we have NAs for any tracts in the final data that aren't
         # included in the data source
-        assert np.all(df[df.MERGE == "left_only"][final_columns].isna())
+        has_additional_non_null_tracts = not np.all(
+            df[df.MERGE == "left_only"][final_columns].isna()
+        )
+        if has_additional_non_null_tracts:
+            # We backfill island areas with data from the 2010 census, so if THOSE tracts
+            # have data beyond the data source, that's to be expected and is fine to pass.
+            # If some other state or territory does though, this should fail
+            left_only = df.loc[(df.MERGE == "left_only")]
+            left_only_has_value = left_only.loc[~df[final_columns].isna().all(axis=1)]
+            fips_with_values = set(left_only_has_value[field_names.GEOID_TRACT_FIELD].str[0:2])
+            non_island_fips_codes = fips_with_values.difference(TILES_ISLAND_AREA_FIPS_CODES)
+            assert not non_island_fips_codes
+
 
         # Make sure the datasource doesn't have a ton of unmatched tracts, implying it
         # has moved to 2020 tracts