1ru.5: Integrate llama.cpp as a first-class local agent runtime

[penso]

Automated handoff for 1ru.5.

Issue:
Base branch: main
Head branch: beads-polyphony-1ru.5
Commit: ff95ce6

[penso]

Summary

This branch adds a new llama.cpp agent runtime and reuses the OpenAI-compatible chat loop for it. The overall shape is reasonable, but I found a few correctness gaps in wiring and config handling that are likely to show up in real use. I reviewed the diff and surrounding code only, I did not run the full Rust test/lint suite.

Risks

• High: the registry registers LlamaRuntime with default() instead of passing the shared ToolExecutor, so tool-capable llama runs will always report unsupported tool calls even though the runtime itself is built to execute them. Compare crates/agents/src/lib.rs:24 and crates/agents/src/lib.rs:48 with crates/agent-llama/src/lib.rs:44 and crates/agent-llama/src/lib.rs:205.
• Medium: kind: llama_cpp is accepted by the runtime’s supports() check, but workflow transport inference does not map that kind and falls back to LocalCli unless the user also sets transport. That makes one supported spelling silently select the wrong backend. See crates/workflow/src/render.rs:406 and crates/agent-llama/src/lib.rs:175.
• Medium: config validation now allows llama profiles with no model as long as fetch_models is true, but the rendered agent still defaults to command: llama-server, and auto-start later hard-fails because spawning requires a model path. That means an invalid local setup is accepted and only breaks at runtime. See crates/workflow/src/service.rs:562, crates/workflow/src/render.rs:437, and crates/agent-llama/src/lib.rs:299.

Recommended human checks

• Exercise a real llama.cpp run that emits a tool call and confirm the event stream contains ToolCallCompleted, not unsupported_tool_call.
• Try three config variants, kind: llama, kind: llama_cpp, and explicit transport: llama_cpp, and verify all three select the new runtime.
• Test both startup modes explicitly: a pre-running remote/local server with no spawn command, and local auto-spawn via llama-server with and without a configured model path.