Studio: report persisted provenance from the dedup reply too

Both review points on the previous revision were correct.

- the already_loaded GGUF response still derived provenance from the
  filesystem, so a deduplicated /load flipped a local model to remote
  once its directory went away, the same flip already fixed for the
  status poll. It now reads the persisted value
- drop the .gguf suffix from the local classification in the settings
  sheet. A one-slash org/name.gguf is a repository id, not a file, as
  _is_direct_gguf_file_ref documents, so that suffix overrode the
  backend saying remote and offered filesystem placement guidance for a
  model that downloads

The contract now pins both GGUF responses using the provenance helper,
and pins the suffix out of the classification.

studio/backend/routes/inference.py

@@ -4509,8 +4509,9 @@ async def _load_model_impl(
                     is_vision = llama_backend._is_vision,
                     is_lora = False,
                     is_gguf = True,
-                    is_local_model = native_grant_backed
-                    or is_local_path(llama_backend.model_identifier),
+                    is_local_model = _loaded_is_local_model(
+                        llama_backend, native_grant_backed, llama_backend.model_identifier
+                    ),
                     is_diffusion = llama_backend.is_diffusion,
                     is_audio = _gguf_is_audio,
                     audio_type = _gguf_audio,

studio/frontend/src/features/chat/chat-settings-sheet.tsx

@@ -394,14 +394,12 @@ export function ChatSettingsPanel({
     ggufContextLength != null ||
     (currentCheckpoint?.toLowerCase().endsWith(".gguf") ?? false);
   // activeModelIsLocal is the backend's own classification and covers native
-  // picks. activeNativePathToken must not be used here: status reconciliation
-  // keeps it across a switch to a remote GGUF (no replacement token exists),
-  // so a stale token would label that remote model local.
+  // picks. Two things must not decide this: activeNativePathToken, which
+  // status reconciliation keeps across a switch to a remote GGUF (no
+  // replacement token exists), and a bare .gguf suffix, since the backend
+  // reads a one-slash org/name.gguf as a repository id, not a file.
   const isLocalGguf =
-    isGguf &&
-    (activeModelIsLocal ||
-      isLocalModelPath(currentCheckpoint ?? "") ||
-      (currentCheckpoint?.toLowerCase().endsWith(".gguf") ?? false));
+    isGguf && (activeModelIsLocal || isLocalModelPath(currentCheckpoint ?? ""));
   const ggufMaxContextLength = useChatRuntimeStore(
     (s) => s.ggufMaxContextLength,
   );

tests/studio/test_model_picker_contracts.py

@@ -351,9 +351,12 @@ def test_local_mtp_warning_covers_path_and_native_gguf_sources():
     assert "isGguf &&" in local.group(0)
     assert "activeModelIsLocal" in local.group(0)
     assert "isLocalModelPath" in local.group(0)
-    # A native token outlives a switch to a remote GGUF, so it must not
-    # classify the model here; activeModelIsLocal already covers native picks.
+    # Two signals must not classify the model here, because both mislabel a
+    # remote GGUF as local: a native token, which outlives a switch to a remote
+    # model, and a bare .gguf suffix, since a one-slash org/name.gguf is a
+    # repository id. activeModelIsLocal is the backend's own answer for both.
     assert "activeNativePathToken" not in local.group(0)
+    assert ".gguf" not in local.group(0)
     assert "isLocalGguf" in src.split('specFallbackReason === "drafter_not_found"', 1)[1]
 
 
@@ -380,7 +383,9 @@ def test_local_mtp_warning_uses_backend_source_metadata():
     # the filesystem would flip a local model to remote once its directory goes
     # away underneath a running server.
     assert "llama_backend._is_local_model = bool(native_grant_backed or config.is_local)" in route
-    assert "is_local_model = _loaded_is_local_model(" in route
+    # Both GGUF responses report it: the status poll and the already_loaded
+    # dedup reply. Either one re-deriving it reintroduces the flip.
+    assert route.count("is_local_model = _loaded_is_local_model(") >= 2
     assert "backend.active_model_name and is_local_path(backend.active_model_name)" in route