Ship portable CPU dispatch and verified shared runtime dependencies

tools/embeddinggemma2/package_smoke.py

@@ -22,6 +22,7 @@ def main():
     parser.add_argument("--archive", type=Path, default=ROOT / ".searchi/packages/searchi-windows.zip")
     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")
     args = parser.parse_args()
     manifest = json.loads(Path(__file__).with_name("model-manifest.json").read_text())
     model = args.models / manifest["files"][0]["path"]
@@ -40,6 +41,8 @@ def main():
                 data = archive.read(record["path"])
                 assert len(data) == record["bytes"]
                 assert hashlib.sha256(data).hexdigest() == record["sha256"]
+                if args.baseline and record["path"].startswith("ggml-cpu-") and record["path"] != "ggml-cpu-x64.dll":
+                    continue
                 path.parent.mkdir(parents=True, exist_ok=True)
                 path.write_bytes(data)
         environment = dict(os.environ, GGML_DISABLE_VULKAN="1")
@@ -56,10 +59,13 @@ def main():
             worker.kill()
             worker.wait(timeout=10)
             raise
-        assert worker.returncode == 0, stderr
+        assert worker.returncode == 0, f"Worker exit {worker.returncode:#x}: {stderr}"
         worker_duration_ms = (time.perf_counter()-started)*1000
         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
+        if args.baseline:
+            assert hello.get("cpu_backend_module") == "ggml-cpu-x64.dll"
         vector = result["vectors"][0]
         assert len(vector) == 768 and all(math.isfinite(x) for x in vector)
         assert abs(sum(x*x for x in vector) - 1) < 0.0001
@@ -69,6 +75,9 @@ def main():
         with args.archive.open("rb") as stream:
             archive_checksum = hashlib.file_digest(stream, "sha256").hexdigest()
         report = {"architecture": package["architecture"], "verified_files": len(package["files"]),
+            "cpu_backend_module": hello.get("cpu_backend_module"),
+            "vulkan_module_loaded": hello.get("vulkan_module_loaded"),
+            "baseline_only": args.baseline,
             "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",