feat(goal): /config skeptic-local — auto-managed local model for the skeptic

[davidrhodus]

Stacked on #110 (opt-in skeptic_endpoint). Makes the local /goal skeptic fully automated behind one toggle: /config skeptic-local on.

What it does

/config skeptic-local on runs the whole setup, each step failing soft:

1. Detect backend — Apple-Silicon MLX or NVIDIA CUDA (nvidia-smi probe). No backend → message, skeptic stays on the main model.
2. Resolve model — a small default review model (~3B, 4-bit) per backend, overridable via HI_SKEPTIC_LOCAL_REPO / HI_SKEPTIC_LOCAL_MODEL_ID / HI_SKEPTIC_LOCAL_GGUF.
3. Download if missing — reuses the same .hi/models cache layout as /hf run --mlx, so a model fetched either way is shared.
4. Launch + health-poll — spawns hi-local serve and waits for /health (verbatim of the proven /hf path).
5. Wire it — sets skeptic_endpoint/skeptic_model and rebuilds the skeptic provider. The driver and planner stay on the main model.

/config skeptic-local off stops the server and restores the prior skeptic settings.

Safety / graceful degradation

• Config is only mutated after the server reports healthy — any failure (no backend, missing hi-local, download fail, unhealthy server) returns an error and leaves the skeptic on the main provider. No hangs, no half-wired state.
• Server is torn down on session exit: CLI via kill_background_processes, TUI via a drop guard covering every exit path (private registry, stop_all_local_servers).
• The blocking, progress-to-terminal download runs only on the plain CLI; the TUI returns NeedsDownload instead of corrupting its alternate screen.

Structure

• hi-tools/local_server.rs (new) — process/HTTP mechanics: spawn into a private background registry, /health wait, stop-by-id / stop-all, skeptic_model_dir cache path.
• hi-agent/local_skeptic.rs (new) — policy + orchestration: backend detection, default-model registry, presence checks, serve arg builder, and Agent::enable_local_skeptic / disable_local_skeptic.
• /config skeptic-local <on|off> parsing (ConfigArg::SkepticLocal) wired into both frontends; the CLI routes it through an async arm (it may download/spawn).

Tests

• 8 unit tests over the deterministic core (backend selection, model registry, presence checks, arg builder, config parsing/routing). cargo test -p hi-agent -p hi-tools green (564 + 229).
• fmt / clippy (-D warnings) / file-size ratchet all pass.
• Not exercised in CI: the live multi-GB download + real MLX/CUDA server spawn — validated on local hardware.

🤖 Generated with Claude Code