refactor: remove embedded llama.cpp, rely on Ollama for local inference

[wmeddie]

Summary

Remove embedded llama.cpp entirely. Local inference goes through Ollama only.

Stacked on #102 — base branch is fix/per-agent-llm-config-v2, not main. Land #102 first, then this.

Why

Two recurring problems with the embedded path:

1. Crashes. llama-cpp-2 invariant violations (KV-cache type mismatches, sampler lifetime issues) brought down the whole server process. Tracking these required deep familiarity with llama.cpp internals and was a tax on every other change in the LLM stack.
2. Hardware-build complexity. Releases needed cross-compiling with the right combination of features per platform (metal for Apple Silicon, cuda for Linux+NVIDIA, plain CPU otherwise). CUDA in particular needed bespoke CUDA_PATH/RUSTFLAGS setup in build.sh because find_cuda_helper only searches <root>/lib64, missing Debian/Ubuntu multiarch layouts.

Ollama solves both: it ships its own platform-tuned builds, runs out-of-process so crashes don't take down the server, and we already speak its HTTP API.

ADR-023 captures the decision; ADR-011 is marked superseded.

What changed

Removed:

• crates/xpressclaw-core/src/llm/llamacpp.rs (~1400 lines)
• llama-cpp-2, hf-hub, encoding_rs workspace dependencies
• local-llm, metal, cuda Cargo features (core, server, cli)
• AgentLlmConfig.model_path field
• The "local" arm in LlmRouter::materialize_provider
• DownloadProgress / DownloadStatus / download_status route
• use_embedded request flag and the wizard's download-progress UI
• resolve_gguf_source and the post-setup GGUF download flow
• nvcc/Metal detection in build.sh
• --skip llamacpp filter in CI

Kept:

• provider: ollama with HTTP proxy through LocalProvider
• The reconciler's per-host Ollama pull loop (reconcile_models from refactor: per-agent LLM config with logical-name routing #102) — this is now the only model-acquisition path
• Hardware detection / recommend_model — feeds the Ollama tag picker

Net: -1948 lines.

Future (not in this PR)

XpressAI has GGUF builds (e.g. Qwen3.6-27B-RYS-UD-Q4_K_XL-GGUF) that outperform base Qwen3.5/3.6. Using them after this PR requires publishing to Ollama Hub first. ADR-023's "Future" section captures it.

A POST /api/setup/pull-model with SSE progress would be a nice follow-up so the wizard can show "pulling X of Y MB" without reintroducing the embedded path.

Migration

Existing xpressai.yaml files using provider: local will warn at router-build time and the agent will be skipped:

unknown provider type 'local'. Supported providers: openai, anthropic, ollama.

Edit the agent's llm block: provider: ollama, set the right model: tag, xpressclaw up. The reconciler pulls the model from Ollama on first start.

Test plan

• cargo fmt --all --check clean
• cargo clippy --workspace --all-targets clean
• cargo test --workspace — 401 passed (down 1 from refactor: per-agent LLM config with logical-name routing #102 because the 6 ignored llamacpp tests are gone with the file)
• npx svelte-check — 0 errors
• Manual: setup wizard with provider=ollama and a model that isn't yet pulled — verify agent appears, reconciler pulls in background, agent starts replying once pull completes
• Manual: edit existing config to provider: local — verify the agent is skipped with the documented error
• Manual: build with ./build.sh and confirm no GPU SDK is required