feat(autocomplete): group relative-path results by folder (#1091)

Autocomplete suggestions were ranked purely by relevance across the whole
library, so same-named loras from different subfolders interleaved and were
hard to tell apart. Results are now bucketed by folder (root first, then
alphabetically, with nested paths sorting naturally) while keeping the
existing relevance ordering within each folder group.

tests/services/test_relative_path_search.py

@@ -49,9 +49,83 @@ async def test_search_relative_paths_supports_multiple_tokens():
 
     matching = await service.search_relative_paths("flux detail")
 
+    # Folder grouping takes precedence over cross-folder relevance:
+    # the "detail" folder sorts before "flux" alphabetically.
     assert matching == [
-        f"flux{os.sep}detail-model.safetensors",
         f"detail{os.sep}flux-trained.safetensors",
+        f"flux{os.sep}detail-model.safetensors",
+    ]
+
+
+@pytest.mark.asyncio
+async def test_search_relative_paths_groups_by_folder_alphabetically():
+    """Same-folder entries cluster together; folders sort alphabetically."""
+    scanner = FakeScanner(
+        [
+            {"file_path": "/models/zeta/model-z1.safetensors"},
+            {"file_path": "/models/alpha/model-a1.safetensors"},
+            {"file_path": "/models/zeta/model-z2.safetensors"},
+            {"file_path": "/models/alpha/model-a2.safetensors"},
+            {"file_path": "/models/model-root.safetensors"},
+        ],
+        ["/models"],
+    )
+    service = DummyService("stub", scanner, BaseModelMetadata)
+
+    matching = await service.search_relative_paths("model")
+
+    assert matching == [
+        # Root-level files (empty folder) come first
+        "model-root.safetensors",
+        f"alpha{os.sep}model-a1.safetensors",
+        f"alpha{os.sep}model-a2.safetensors",
+        f"zeta{os.sep}model-z1.safetensors",
+        f"zeta{os.sep}model-z2.safetensors",
+    ]
+
+
+@pytest.mark.asyncio
+async def test_search_relative_paths_relevance_within_folder_group():
+    """Within a folder group, the relevance ordering still applies."""
+    scanner = FakeScanner(
+        [
+            {"file_path": "/models/flux/x-detail-model.safetensors"},
+            {"file_path": "/models/flux/detail-model.safetensors"},
+            {"file_path": "/models/flux/a-very-long-detail-model-name.safetensors"},
+        ],
+        ["/models"],
+    )
+    service = DummyService("stub", scanner, BaseModelMetadata)
+
+    matching = await service.search_relative_paths("flux detail")
+
+    assert matching == [
+        # Prefix hit on the full path wins
+        f"flux{os.sep}detail-model.safetensors",
+        # Then earliest match position, then shorter path
+        f"flux{os.sep}x-detail-model.safetensors",
+        f"flux{os.sep}a-very-long-detail-model-name.safetensors",
+    ]
+
+
+@pytest.mark.asyncio
+async def test_search_relative_paths_nested_folders_sort_naturally():
+    scanner = FakeScanner(
+        [
+            {"file_path": "/models/styles/anime/model-b.safetensors"},
+            {"file_path": "/models/styles/model-a.safetensors"},
+            {"file_path": "/models/other/model-c.safetensors"},
+        ],
+        ["/models"],
+    )
+    service = DummyService("stub", scanner, BaseModelMetadata)
+
+    matching = await service.search_relative_paths("model")
+
+    assert matching == [
+        f"other{os.sep}model-c.safetensors",
+        f"styles{os.sep}model-a.safetensors",
+        f"styles{os.sep}anime{os.sep}model-b.safetensors",
     ]