server: add new exception types for 400 responses

CMakeLists.txt

@@ -3095,6 +3095,33 @@ if (ENGINE_BUILD_TESTS OR ENGINE_BUILD_EXTENDED_TESTS OR ENGINE_BUILD_MODEL_TEST
         add_test(NAME server_parallel_lifecycle_test COMMAND server_parallel_lifecycle_test)
         set_tests_properties(server_parallel_lifecycle_test PROPERTIES TIMEOUT 90)
 
+        # Exercise the server exception conversion to HTTP responses
+        get_target_property(server_exception_sources audiocpp_server SOURCES)
+        list(FILTER server_exception_sources EXCLUDE REGEX "app/server/(main|model_installer|parallel_runtime)\\.cpp$")
+        add_executable(server_request_exceptions_test
+            tests/unittests/test_server_request_exception.cpp ${server_exception_sources})
+        target_include_directories(server_request_exceptions_test PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}/app/server "${CMAKE_CURRENT_BINARY_DIR}/generated")
+        target_link_libraries(server_request_exceptions_test PRIVATE engine_runtime ggml Threads::Threads)
+        if(WIN32)
+            target_link_libraries(server_request_exceptions_test PRIVATE ws2_32)
+        endif()
+        add_test(NAME server_request_exceptions_test COMMAND server_request_exceptions_test)
+        set_tests_properties(server_request_exceptions_test PROPERTIES TIMEOUT 60)
+
+        get_target_property(server_parallel_exception_sources audiocpp_server SOURCES)
+        list(FILTER server_parallel_exception_sources EXCLUDE REGEX "app/server/(main|model_installer|runtime)\\.cpp$")
+        add_executable(server_parallel_request_exceptions_test
+            tests/unittests/test_server_parallel_request_exception.cpp ${server_parallel_exception_sources})
+        target_include_directories(server_parallel_request_exceptions_test PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}/app/server "${CMAKE_CURRENT_BINARY_DIR}/generated")
+        target_link_libraries(server_parallel_request_exceptions_test PRIVATE engine_runtime ggml Threads::Threads)
+        if(WIN32)
+            target_link_libraries(server_parallel_request_exceptions_test PRIVATE ws2_32)
+        endif()
+        add_test(NAME server_parallel_request_exceptions_test COMMAND server_parallel_request_exceptions_test)
+        set_tests_properties(server_parallel_request_exceptions_test PROPERTIES TIMEOUT 60)
+
         add_executable(session_pool_test tests/unittests/test_session_pool.cpp)
         target_include_directories(session_pool_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include ${CMAKE_CURRENT_SOURCE_DIR}/app/server)
         # TaskResult owns JSON values for custom-schema outputs.

app/server/parallel_runtime.cpp

@@ -1283,8 +1283,8 @@ HttpResponse ParallelServerState::handle_request(const HttpRequest & request, bo
     else {
         response = error_response(404, "unknown endpoint: " + request.path, "not_found");
     }
-  } catch (const engine::runtime::CapacityError & ex) {
-    // The request is too big for the device, which is the caller's to fix --
+  } catch (const engine::runtime::RequestValidationError & ex) {
+    // Something is wrong with the request that the caller can fix --
     // reporting it as 500 sends them looking for a server fault that is not
     // there. Checked before ServerBusyError only because both are
     // runtime_error; the two conditions are disjoint.

app/server/runtime.cpp

@@ -1274,8 +1274,8 @@ HttpResponse ServerState::handle_request(const HttpRequest & request, bool use_f
     else {
         response = error_response(404, "unknown endpoint: " + request.path, "not_found");
     }
-  } catch (const engine::runtime::CapacityError & ex) {
-    // The request is too big for the device, which is the caller's to fix --
+  } catch (const engine::runtime::RequestValidationError & ex) {
+    // Something is wrong with the request that the caller can fix --
     // reporting it as 500 sends them looking for a server fault that is not
     // there. Checked before ServerBusyError only because both are
     // runtime_error; the two conditions are disjoint.
@@ -1939,7 +1939,7 @@ void ServerState::ensure_model_loaded_locked(LoadedModel & model) {
         // That is acceptable: the 503 tells the client to retry later.
     }
     ensure_model_fits_memory(model.config);
-    auto registry = engine::runtime::make_default_registry();
+    auto registry = registry_factory_();
 
     engine::runtime::ModelLoadRequest load_request;
     load_request.model_path = model.config.path;

app/server/runtime.h

@@ -13,6 +13,7 @@
 #include "engine/framework/io/json.h"
 #include "engine/framework/runtime/model.h"
 #include "engine/framework/runtime/session.h"
+#include "engine/framework/runtime/registry.h"
 
 #include <atomic>
 #include <cstdint>
@@ -92,6 +93,12 @@ class ServerState final : public IHttpHandler, public ServerFrontendContext {
         void unload();
     };
 
+    // Controlled runtime tests suppy loaders and barriers without production flags.
+    friend class ServerRuntimeTestAccess;
+    std::function<engine::runtime::ModelRegistry()> registry_factory_ = [] {
+        return engine::runtime::make_default_registry();
+    };
+
     // Acquire the model's run guard. `request_timeout_ms` is the caller-supplied
     // override, clamped by this model's configured ceiling. Throws ServerBusyError
     // (-> HTTP 503) once the effective timeout has elapsed.

include/engine/framework/runtime/errors.h

@@ -5,13 +5,19 @@
 
 namespace engine::runtime {
 
+// A request has a problem that the client can fix by changing request
+// parameters. Distinct from a genuine internal fault: so servers should
+// surface it as a client error rather than an opaque 500.
+class RequestValidationError : public std::runtime_error {
+public:
+    explicit RequestValidationError(const std::string & message) : std::runtime_error(message) {}
+};
+
 // A request the device cannot serve AT THIS SIZE -- e.g. a transcription
-// prompt plus audio whose prefill graph does not fit in VRAM. Distinct from a
-// genuine internal fault: the caller can fix it by sending less, so servers
-// should surface it as a client error rather than an opaque 500.
-class CapacityError : public std::runtime_error {
+// prompt plus audio whose prefill graph does not fit in VRAM.
+class CapacityError : public RequestValidationError {
 public:
-    explicit CapacityError(const std::string & message) : std::runtime_error(message) {}
+    explicit CapacityError(const std::string & message) : RequestValidationError(message) {}
 };
 
 }  // namespace engine::runtime