minor changes:

* Changed the tests to use an in-module fixture instead of the
  `test_ontology.owl` graph and removed the `test_ontology.owl` file. The
  problem with arbitrary test data is that it is hard to evolve. If possible
  test data should be kept as terse as possible. In this case I just created a
  namespace manager with three namespaces added, of which only two will be
  effective because of <#2077>.
* Added a warning to `NamespaceManager.curie` to warn users that it is not a
  pure function and explain the specific side effects.

rdflib/namespace/__init__.py

@@ -491,10 +491,27 @@ def qname(self, uri: str) -> str:
             return ":".join((prefix, name))
 
     def curie(self, uri: str) -> str:
-        """From a URI, generate a valid CURIE.
+        """
+        From a URI, generate a valid CURIE.
+
+        Result is guaranteed to contain a colon separating the prefix from the
+        name, even if the prefix is an empty string.
+
+        .. warning::
+
+            If there is no matching namespace for the URI in the namespace
+            manager then a new namespace will be added with prefix
+            ``ns{index}``.
+
+            Because of this side effect this is not a pure function.
+
+            This is the same behaviour as `NamespaceManager.qname`.
+
+
+
+        :param uri: URI to generate CURIE for.
+        :return: CURIE for the URI.
 
-        Result is guaranteed to contain a colon separating the prefix from the name,
-        even if the prefix is an empty string.
         """
         prefix, namespace, name = self.compute_qname(uri)
         return ":".join((prefix, name))