Preserve Claude model and thinking preferences (#2912)

* fix(agents): preserve Claude thinking selections

Model switches now resolve thinking against the target model's saved preference instead of resetting to the catalog default. Claude owns its configured-model defaults and legacy Fable alias handling, while persisted permission modes remain unchanged.

* chore(claude): tag Fable compatibility shim

* fix(app): resolve persisted model aliases

* fix(models): preserve aliases for older clients

* fix(models): hide compatibility aliases from listings

* fix(settings): hide compatibility model aliases

* test(claude): cover legacy catalog entry

* test(app): expect mock thinking controls in schedules

* fix(app): preserve queued thinking preferences

* fix(app): reconcile schedule provider snapshots

* fix(app): preserve schedule model selections

* fix(claude): preserve exact configured model ids

* fix(app): preserve committed thinking preferences

* fix(app): preserve aliased schedule thinking

packages/app/src/provider-selection/provider-selection.test.ts

@@ -66,6 +66,24 @@ describe("combined model selector data", () => {
     ]);
   });
 
+  it("hides compatibility-only model entries from new clients", () => {
+    const compatibilityModel: AgentModelDefinition = {
+      ...codexModel,
+      id: "gpt-5.4-legacy",
+      label: "GPT-5.4 legacy",
+      isSelectable: false,
+    };
+
+    const [provider] = buildSelectableProviderSelectorProviders([
+      snapshotEntry({ provider: "codex", models: [codexModel, compatibilityModel] }),
+    ]);
+
+    expect(provider?.modelSelection).toMatchObject({
+      kind: "models",
+      rows: [{ modelId: "gpt-5.4" }],
+    });
+  });
+
   it("synthesizes a default model row for ready enabled providers without explicit models", () => {
     expect(
       buildSelectableProviderSelectorProviders([