chore(test-project): drop hardcoded openai/gpt-4o so env takes precedence

The embedded llm block overrode ~/.inkos/.env, breaking bench runs that
relied on INKOS_LLM_MODEL / INKOS_LLM_BASE_URL. Removing it lets the
CLI resolve the provider from the user env as documented.

test-project/inkos.json

@@ -1,12 +1,6 @@
 {
   "name": "test-project",
   "version": "0.1.0",
-  "llm": {
-    "provider": "openai",
-    "baseUrl": "https://api.openai.com/v1",
-    "model": "gpt-4o",
-    "stream": false
-  },
   "notify": [],
   "daemon": {
     "schedule": {
@@ -16,4 +10,4 @@
     "maxConcurrentBooks": 3
   },
   "language": "zh"
-}
+}