feat(cli): add jarvis model convert for local HF→MLX/GGUF conversion

[gilbert-barajas]

Summary

Adds jarvis model convert <hf_repo> — downloads raw HuggingFace weights and converts them to the active engine's format, closing the gap for models that exist on HF but ship no pre-built GGUF or MLX artifact. Today jarvis model pull only fetches pre-converted artifacts (Ollama /api/pull, a specific .gguf, or a pre-quantized mlx_repo); if a model has none, there's no path forward. This wires up the well-tested conversion tools — mlx_lm.convert for Apple MLX and convert_hf_to_gguf.py + llama-quantize for llama.cpp.

Closes #251.

What it does

jarvis model convert <hf_repo> [--engine ENGINE] [--quantize q4_k_m] [-q/--mlx-4bit] [--output DIR] [--force]

• MLX (--engine mlx): mlx_lm.convert(...), with -q for 4-bit. Guarded so a missing inference-mlx extra prints an install hint instead of a stack trace.
• GGUF (--engine llamacpp): shells convert_hf_to_gguf.py, then llama-quantize when --quantize is given. Tools located via $LLAMA_CPP_DIR or PATH, with graceful "not found" hints (mirroring how pull's hf_download handles a missing huggingface-cli).
• Engine resolution mirrors pull (--engine → config.engine.default → mlx). Non-convertible engines (ollama/vllm/sglang) print an informational note pointing at pull (exit 0).
• Refuses to overwrite a non-empty --output without --force. On success, prints the artifact path and the exact command to serve it.

It's a subcommand on the existing model group in cli/model.py, following pull's conventions (Click, rich, sys.exit(1) on error).

Testing

tests/cli/test_model_convert.py — 7 CliRunner tests with converters fully mocked (mlx_lm.convert, subprocess.run, hf_download), so nothing downloads or shells in the default lane:

• mlx + -q calls mlx_lm.convert(quantize=True)
• missing inference-mlx → install hint, exit 1
• engine-default resolution (no --engine → config default)
• gguf shells the converter + llama-quantize
• missing tool → install hint, exit 1
• non-empty --output refusal + --force override
• ollama → informational, exit 0

Gates run locally with the pinned toolchain:

• uv run ruff check src/ tests/ — clean
• uv run ruff format --check src/ tests/ — clean
• uv run pytest tests/cli/test_model_convert.py -v — 7 passed

Out of scope (happy to follow up)

Auto-registering the converted model into the catalog so model list surfaces it without the engine running — left out to keep this focused; glad to add it in a follow-up if useful.

🤖 Generated with Claude Code

[josephdhand568-cpu]

Open Jarvis