Check packaged query and indexing contexts during CPU fallback

tools/embeddinggemma2/package_smoke.py

@@ -23,6 +23,7 @@ def main():
     parser.add_argument("--models", type=Path, default=ROOT / ".searchi/models/embeddinggemma2")
     parser.add_argument("--output", type=Path, default=ROOT / ".searchi/packages/smoke.json")
     parser.add_argument("--baseline", action="store_true", help="Extract only the baseline CPU module to test the packaged fallback")
+    parser.add_argument("--role", choices=("indexing", "query"), default="indexing")
     args = parser.parse_args()
     manifest = json.loads(Path(__file__).with_name("model-manifest.json").read_text())
     model = args.models / manifest["files"][0]["path"]
@@ -49,7 +50,7 @@ def main():
         environment.pop("SEARCHI_FORCE_CPU", None)
         started = time.perf_counter()
         worker = subprocess.Popen([str(target / "searchi-gemma-worker.exe"), "--model", str(model),
-            "--threads", "1", "--device", "auto"], stdin=subprocess.PIPE, stdout=subprocess.PIPE,
+            "--threads", "1", "--device", "auto", "--role", args.role], stdin=subprocess.PIPE, stdout=subprocess.PIPE,
             stderr=subprocess.PIPE, text=True, encoding="utf-8", env=environment,
             creationflags=subprocess.CREATE_NO_WINDOW)
         request = json.dumps({"op": "code_query", "text": "retry failed requests with exponential backoff"}) + "\n"
@@ -64,6 +65,8 @@ def main():
         hello, result = [json.loads(line) for line in stdout.splitlines()]
         assert hello["ready"] and hello["device"] == "cpu" and result["device"] == "cpu" and result["ok"]
         assert hello.get("vulkan_module_loaded") is False
+        assert hello.get("role") == args.role
+        assert hello.get("context_tokens") == (512 if args.role == "query" else 2048)
         if args.baseline:
             assert hello.get("cpu_backend_module") == "ggml-cpu-x64.dll"
         vector = result["vectors"][0]
@@ -78,6 +81,7 @@ def main():
             "cpu_backend_module": hello.get("cpu_backend_module"),
             "vulkan_module_loaded": hello.get("vulkan_module_loaded"),
             "baseline_only": args.baseline,
+            "worker_role": hello["role"], "context_tokens": hello["context_tokens"],
             "archive_sha256": archive_checksum,
             "worker_requested_device": "auto", "worker_actual_device": hello["device"],
             "no_gpu_condition": "GGML_DISABLE_VULKAN=1 only in child; SEARCHI_FORCE_CPU absent",