Ship portable CPU dispatch and verified shared runtime dependencies

crates/searchi-semantic/native/worker.cpp

@@ -1,6 +1,7 @@
 // Local NDJSON helper around the pinned llama.cpp encoder. stdout is exclusively
 // protocol output; llama diagnostics go to stderr. No model download or HTTP.
 #include "llama.h"
+#include "ggml-backend.h"
 #include "mtmd.h"
 #include "nlohmann/json.hpp"
 #include <algorithm>
@@ -14,6 +15,12 @@
 #include <stdexcept>
 #include <string>
 #include <vector>
+#ifdef SEARCHI_CPU_DISPATCH
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#include <filesystem>
+#endif
 
 using json = nlohmann::json;
 constexpr size_t max_request_bytes = 16 * 1024 * 1024;
@@ -71,6 +78,8 @@ class Encoder {
     }
 public:
     std::string device;
+    std::string cpu_backend_module;
+    bool vulkan_module_loaded = false;
     int threads;
     Encoder(const std::string & path, int cpu_threads, bool want_gpu) : model_path(path),threads(cpu_threads) {
         llama_log_set([](ggml_log_level level, const char * text, void *) {
@@ -79,6 +88,31 @@ public:
         mtmd_log_set([](ggml_log_level level, const char * text, void *) {
             if (level >= GGML_LOG_LEVEL_WARN) std::fputs(text,stderr);
         },nullptr);
+#ifdef SEARCHI_CPU_DISPATCH
+        // Only load our adjacent modules: no current-directory or environment
+        // plugin discovery. CPU-only startup never loads the Vulkan module.
+        std::vector<wchar_t> executable(32768);
+        DWORD length = GetModuleFileNameW(nullptr, executable.data(), DWORD(executable.size()));
+        require(length > 0 && length < executable.size(), "Cannot locate packaged worker backends");
+        const auto directory = std::filesystem::path(std::wstring(executable.data(), length)).parent_path().u8string();
+        if (want_gpu) {
+            ggml_backend_load_all_from_path(directory.c_str());
+        } else {
+            ggml_backend_load_cpu_only_from_path(directory.c_str());
+        }
+        require(ggml_backend_reg_by_name("CPU") != nullptr, "Cannot load compatible packaged CPU backend");
+        const auto cpu_reg = ggml_backend_reg_by_name("CPU");
+        const auto feature_proc = ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_get_features");
+        HMODULE cpu_module = nullptr;
+        if (feature_proc && GetModuleHandleExW(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT,
+                reinterpret_cast<LPCWSTR>(feature_proc), &cpu_module)) {
+            DWORD module_length = GetModuleFileNameW(cpu_module, executable.data(), DWORD(executable.size()));
+            if (module_length > 0 && module_length < executable.size()) {
+                cpu_backend_module = std::filesystem::path(std::wstring(executable.data(), module_length)).filename().u8string();
+            }
+        }
+        vulkan_module_loaded = GetModuleHandleW(L"ggml-vulkan.dll") != nullptr;
+#endif
         llama_backend_init();
         bool use_gpu = want_gpu && llama_supports_gpu_offload();
         require(open(use_gpu) || (use_gpu && open(false)),"Cannot initialize local EmbeddingGemma 2 model/context");
@@ -335,14 +369,17 @@ int main(int argc, char ** argv) {
         // probing/enabling the Vulkan provider in the Windows VM.
         if (force_cpu || device == "cpu") {
 #ifdef _WIN32
+            require(_putenv_s("GGML_CPU_ONLY", "1") == 0, "Cannot restrict CPU startup");
             require(_putenv_s("GGML_DISABLE_VULKAN", "1") == 0, "Cannot disable Vulkan for CPU startup");
 #else
+            require(setenv("GGML_CPU_ONLY", "1", 1) == 0, "Cannot restrict CPU startup");
             require(setenv("GGML_DISABLE_VULKAN", "1", 1) == 0, "Cannot disable Vulkan for CPU startup");
 #endif
         }
         Encoder encoder(path,threads,device == "auto" && !force_cpu);
         std::cout << json({{"ready",true},{"model","google/embeddinggemma-2"},{"dimensions",dimensions},
-            {"device",encoder.device},{"threads",threads},{"runtime",SEARCHI_LLAMA_TAG},{"media_supported",true},{"code_query_supported",true}}).dump() << '\n' << std::flush;
+            {"device",encoder.device},{"threads",threads},{"runtime",SEARCHI_LLAMA_TAG},{"media_supported",true},{"code_query_supported",true},
+            {"cpu_backend_module",encoder.cpu_backend_module},{"vulkan_module_loaded",encoder.vulkan_module_loaded}}).dump() << '\n' << std::flush;
         for (;;) {
             std::string line; char c;
             while (std::cin.get(c) && c != '\n') {