bdd tests: do not query word table directly

The BDD tests cannot make assumptions about the structure of the
word table anymore because it depends on the tokenizer. Use more
abstract descriptions instead that ask for specific kinds of
tokens.

test/bdd/steps/steps_db_ops.py

@@ -281,6 +281,39 @@ def check_word_table(context, exclude):
             else:
                 assert cur.rowcount > 0, "Row not in word table: %s" % '/'.join(values)
 
+
+@then("there are(?P<exclude> no)? word tokens for postcodes (?P<postcodes>.*)")
+def check_word_table_for_postcodes(context, exclude, postcodes):
+    """ Check that the tokenizer produces postcode tokens for the given
+        postcodes. The postcodes are a comma-separated list of postcodes.
+        Whitespace matters.
+    """
+    nctx = context.nominatim
+    tokenizer = tokenizer_factory.get_tokenizer_for_db(nctx.get_test_config())
+    with tokenizer.name_analyzer() as ana:
+        plist = [ana.normalize_postcode(p) for p in postcodes.split(',')]
+
+    plist.sort()
+
+    with context.db.cursor(cursor_factory=psycopg2.extras.DictCursor) as cur:
+        if nctx.tokenizer == 'legacy_icu':
+            cur.execute("""SELECT info->>'postcode' FROM word
+                           WHERE type = 'P' and info->>'postcode' = any(%s)""",
+                        (plist,))
+        else:
+            cur.execute("""SELECT word FROM word WHERE word = any(%s)
+                             and class = 'place' and type = 'postcode'""",
+                        (plist,))
+
+        found = [row[0] for row in cur]
+        assert len(found) == len(set(found)), f"Duplicate rows for postcodes: {found}"
+
+    if exclude:
+        assert len(found) == 0, f"Unexpected postcodes: {found}"
+    else:
+        assert set(found) == set(plist), \
+        f"Missing postcodes {set(plist) - set(found)}. Found: {found}"
+
 @then("place_addressline contains")
 def check_place_addressline(context):
     """ Check the contents of the place_addressline table. Each row represents