Add local spoken passages with Google retrieval and isolated runtime setup

tools/speech-runtime/cpu-only-loader.patch

@@ -0,0 +1,31 @@
+--- a/ggml/include/ggml-backend.h
++++ b/ggml/include/ggml-backend.h
+@@ -260,6 +260,8 @@
+     // Load all known backends from dynamic libraries
+     GGML_API void               ggml_backend_load_all(void);
+     GGML_API void               ggml_backend_load_all_from_path(const char * dir_path);
++    // Searchi: explicit CPU-only startup must never load or probe GPU modules.
++    GGML_API void               ggml_backend_load_cpu_only_from_path(const char * dir_path);
+ 
+     //
+     // Backend scheduler
+--- a/ggml/src/ggml-backend-reg.cpp
++++ b/ggml/src/ggml-backend-reg.cpp
+@@ -575,6 +575,17 @@
+     ggml_backend_load_all_from_path(nullptr);
+ }
+ 
++void ggml_backend_load_cpu_only_from_path(const char * dir_path) {
++#ifdef NDEBUG
++    const bool silent = true;
++#else
++    const bool silent = false;
++#endif
++    // Only CPU variants in the owned executable directory. No GPU registry
++    // probes, working-directory discovery or GGML_BACKEND_PATH injection.
++    ggml_backend_load_best("cpu", silent, dir_path);
++}
++
+ void ggml_backend_load_all_from_path(const char * dir_path) {
+ #ifdef NDEBUG
+     bool silent = true;