Package portable Linux runtime and verify Google CPU search

crates/searchi-semantic/native/worker.cpp

@@ -16,10 +16,17 @@
 #include <string>
 #include <vector>
 #ifdef SEARCHI_CPU_DISPATCH
+#include <filesystem>
+#ifdef _WIN32
 #define WIN32_LEAN_AND_MEAN
 #define NOMINMAX
 #include <windows.h>
-#include <filesystem>
+#elif defined(__linux__)
+#include <dlfcn.h>
+#include <unistd.h>
+#else
+#error Unsupported platform for packaged CPU dispatch
+#endif
 #endif
 
 using json = nlohmann::json;
@@ -91,10 +98,19 @@ public:
 #ifdef SEARCHI_CPU_DISPATCH
         // Only load our adjacent modules: no current-directory or environment
         // plugin discovery. CPU-only startup never loads the Vulkan module.
+#ifdef _WIN32
         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();
+#else
+        // /proc identifies this executable even when launched through PATH or
+        // from another working directory. Reject truncation before discovery.
+        std::vector<char> executable(32768);
+        const auto length = readlink("/proc/self/exe", executable.data(), executable.size());
+        require(length > 0 && size_t(length) < executable.size(), "Cannot locate packaged worker backends");
+        const auto directory = std::filesystem::path(std::string(executable.data(), size_t(length))).parent_path().string();
+#endif
         if (want_gpu) {
             ggml_backend_load_all_from_path(directory.c_str());
         } else {
@@ -103,6 +119,7 @@ public:
         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");
+#ifdef _WIN32
         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)) {
@@ -112,6 +129,19 @@ public:
             }
         }
         vulkan_module_loaded = GetModuleHandleW(L"ggml-vulkan.dll") != nullptr;
+#else
+        Dl_info cpu_module{};
+        if (feature_proc && dladdr(feature_proc, &cpu_module) && cpu_module.dli_fname) {
+            cpu_backend_module = std::filesystem::path(cpu_module.dli_fname).filename().string();
+        }
+        // RTLD_NOLOAD observes module presence without initiating provider
+        // loading; forced CPU startup must remain independent of Vulkan.
+        const auto vulkan_path = std::filesystem::path(directory) / "libggml-vulkan.so";
+        if (void * module = dlopen(vulkan_path.c_str(), RTLD_NOW | RTLD_NOLOAD)) {
+            vulkan_module_loaded = true;
+            dlclose(module);
+        }
+#endif
 #endif
         llama_backend_init();
         bool use_gpu = want_gpu && llama_supports_gpu_offload();