Add protocol-neutral model configuration and runtime ownership against the pinned audio.cpp interfaces. Validate task capabilities, preserve the active model on replacement failures, serialize inference calls, and guarantee session-first teardown.
Assisted-by: Codex:gpt-5
Pin and fetch audio.cpp, generate the LocalAI gRPC protocol sources, and propagate the upstream accelerator switches into the engine_runtime-linked server target.
Add a fixture-only contract test covering CPU, CUDA, Vulkan, Metal, legacy option rejection, and uname-based Darwin selection.
Assisted-by: Codex:gpt-5
