llama.cpp fixes XML tool-call ordering

This is exactly the kind of unglamorous parser fix that makes agent stacks feel less brittle in real use. Local inference frameworks now win or lose on tool-call reliability as much as raw model quality.

• –The underlying issue was argument-order rigidity: if a model emitted `limit` before `offset`, the grammar could reject the call even when the intent was correct
• –The failure showed up most clearly on file-reading tools with multiple optional params, where the assistant would retry broken calls and fall into loops
• –Because the change lands in llama.cpp's autoparser layer, downstream apps using its server-side tool calling should benefit without redesigning schemas
• –It is a useful reminder that practical agent UX depends on protocol tolerance and parser ergonomics, not just better benchmarks