[Lorchie]

Fixes #386.

Problem

has_nvidia_gpu() only checked that nvidia-smi was on PATH. On a machine where nvidia-smi.exe is present but no NVIDIA driver/GPU answers (the reporter has an RX 9070 XT), _pick_assets() installed the Windows CUDA build of llama.cpp, so llama-server could not use the GPU and ran on CPU.

Users who already got the wrong build were stuck: the Agent settings only show "Install engine" while the engine is missing.

Changes

• Detection (api/services/llm_server.py): has_nvidia_gpu() now runs nvidia-smi -L and requires a zero exit code plus at least one GPU N: line. Cached, since /llm/status calls it on every poll. No usable NVIDIA GPU → Vulkan build, as before.
• Repair (GET /llm/binary/install?repair=true, Settings › Agent › Engine › Repair): re-detects the GPU, downloads the build, extracts it into bin.new, stops the loaded llama-servers, then swaps bin → bin.old, bin.new → bin. The current engine is only touched once the new one is fully extracted, so a failed or cancelled download keeps a working install, and no stale backend DLL (ggml-cuda.dll) survives the swap.

Testing

• api/tests/test_llm_gpu_detection.py (11 tests): nvidia-smi missing / failing without driver / listing no GPU / crashing / working; Vulkan vs CUDA+cudart selection; repair replaces the old files, keeps them on failed download and on a locked folder, plain install untouched.
• Real run on Windows (RTX 5070) into a temp MODLY_LLM_DIR seeded with a stale DLL: release b10075, CUDA 12.4 + cudart picked, stale DLL gone, llama-server --list-devices sees the GPU. The same release resolves to llama-b10075-bin-win-vulkan-x64.zip when no NVIDIA GPU is detected.
• Also checked on Windows: renaming bin works even while an exe inside it is running, so the swap cannot leave a half-deleted engine; the old copy stays in bin.old and is removed on the next repair.
• npm test (pre-push hook) passes. tsc adds no error, eslint clean.

Also in this PR: two mesh-ops tests that failed on a clean dev

Separate commit, they blocked the pre-push hook:

• test_auto_orient_does_not_flip_a_mesh_over: trimesh's update_faces keeps the cut-away vertices, so the exported bowl still carried the whole sphere and the test measured the bowl's "center" on the orphan top cap. The fixture now drops unreferenced vertices; auto_orient itself was correct.
• test_mesh_ops_run_without_undeclared_networkx runs repair_mesh, which needs pymeshlab: it is now skipped when pymeshlab is missing, like the other repair/inspection tests (all 42 mesh-ops tests pass with pymeshlab installed).
• Not tested: the Repair button clicked in the packaged app, and an actual AMD machine.