Compare the derived tag vocabulary as the database stores tag names

Stored tag names reach the database through normalize_tags, which strips
surrounding whitespace, but the vocabulary was built straight from registered
category names. A category registered as "vae " produced the entry
"model_type:vae " while the stored tag is "model_type:vae", so the entry never
matched and the stale automatic tag it exists to authorise removing survived
the reprojection.

app/assets/services/path_utils.py

@@ -3,6 +3,7 @@ from pathlib import Path
 from typing import Literal
 
 import folder_paths
+from app.assets.helpers import normalize_tags
 
 
 _NON_MODEL_FOLDER_NAMES = frozenset({"configs", "custom_nodes"})
@@ -331,7 +332,10 @@ def get_path_derived_tag_vocabulary() -> set[str]:
     vocabulary.update(_KNOWN_SUBFOLDER_TAGS)
     for folder_name, _bases, _extensions in get_comfy_models_folders():
         vocabulary.add(f"model_type:{folder_name}")
-    return vocabulary
+    # Stored tag names reach the database through normalize_tags, so an
+    # un-normalized vocabulary entry would never match one and the stale tag
+    # it is meant to authorise removing would survive.
+    return set(normalize_tags(list(vocabulary)))
 
 
 def get_name_and_tags_from_asset_path(file_path: str) -> tuple[str, list[str]]:

tests-unit/assets_test/test_semantics_reset.py

@@ -381,6 +381,32 @@ class TestTagReprojection:
         )
         assert "model_type:checkpoints" in _tags(session, "ref-1")
 
+    def test_vocabulary_matches_tags_as_the_database_stores_them(
+        self, session, comfy_dirs
+    ):
+        """Stored names arrive normalized, so a raw vocabulary entry never matches."""
+        path = _write(comfy_dirs["checkpoints"], "model.safetensors")
+        _register(
+            session,
+            path,
+            "ref-1",
+            tags={"model_type:loras": "automatic"},
+        )
+
+        with patch(
+            "app.assets.services.path_utils.get_comfy_models_folders",
+            return_value=[
+                ("checkpoints", [str(comfy_dirs["checkpoints"])], {".safetensors"}),
+                ("loras ", [str(comfy_dirs["loras"])], {".safetensors"}),
+            ],
+        ):
+            reproject_derived_state()
+
+        assert "model_type:loras" not in _tags(session, "ref-1"), (
+            "a category name with stray whitespace must not smuggle an entry "
+            "past the vocabulary and leave the stale tag in place"
+        )
+
     def test_automatic_tag_outside_the_vocabulary_is_left_alone(
         self, session, comfy_dirs
     ):