Let Lemonade auto-select its llama.cpp backend

[rominf]

Summary

• The Lemonade adapter hard-forced lemonade load --llamacpp rocm with no fallback. On GPUs the bundled ROCm build does not support (e.g. the Radeon 780M / gfx1103 iGPU), Lemonade reports llamacpp:rocm as "Unsupported GPU", so every managed load returned a body-less 500 and spun in a load/evict/unload loop. Standalone Lemonade serves the same model because it falls back to Vulkan.
• Query lemonade backends, pick the best supported llama.cpp backend (rocm > vulkan > cpu), install it on demand, and load with it instead of hardcoding rocm.
• Backend is resolved at install and re-resolved at serve time, so existing installs self-heal without a reinstall. Health verification, persisted state, and adapter messaging now follow the selected backend rather than assuming rocm.

Changes (engines/lemonade/src/lib.rs)

• run_lemonade_backends_list + parse_llamacpp_backend_statuses parse the lemonade backends table for the llamacpp recipe.
• select_best_llamacpp_backend chooses the highest-priority supported backend (LLAMACPP_BACKEND_PRIORITY).
• ensure_best_llamacpp_backend installs the chosen backend if needed; used by both install (install_best_llamacpp_backend) and serve_http.
• lemonade_model_load_args / run_lemonade_model_load pass the selected backend; --force dropped from backend install so only supported backends install.
• Health check (health_has_loaded_model / lemonade_backend_matches), state fields, and detect/install/resolve copy reflect the selected backend.

Test plan

• cargo test -p rocm-engine-lemonade (15 passing, incl. new parser/selection tests)
• cargo build --release -p rocm
• Manual: restarted a previously-500-looping Qwen3-4B service; log shows Installing backend: llamacpp:vulkan, Using LlamaCpp Backend: vulkan, model loaded on AMD Radeon 780M Graphics, chat smoke 200 OK (~10.5 tok/s); services list reports ready.

🤖 Generated with Claude Code