llama.cpp router mode hits tool parsing issues

The shift to a native model router is a major leap for llama.cpp, but it exposes the fragile state of local LLM tool-calling compatibility. Tool call parsing errors likely stem from llama-server returning JSON objects for arguments instead of the legacy stringified JSON expected by OpenAI-compatible clients. Ollama's "just works" experience is powered by an internal normalization layer that abstracts away the complexity of divergent model templates. This fragmentation between model quants and engines creates a compatibility tax for raw users, though the new LRU dynamic loading feature remains a significant optimization for multi-model pipelines on consumer VRAM.