import CLI surface

What the mi binary (the CLI entry point of the import crate) exposes today.

Subcommands

mi <SUBCOMMAND>

mi import <DIR> contract

Input: a directory containing

• exactly one *.gguf file
• a tokenizer.json
• a config.json

Output: ~/llm/<NAME>.model, where NAME is derived by stripping -import suffix from the source directory name.

Side effects:

• Reads ~/llm/ to find target path; creates if missing.
• Writes a single .model file. No staging, no atomic rename.
• Embeds the binary IR graph as a ~~~graph section when the config parses as a LlamaStyle family — see graph.md.

The .model invariants are enforced by import.md.

mi download <REPO> contract

Fetch a model from HuggingFace into the local hf-hub cache. The contract — artifact priority, sibling metadata, failure modes — lives in hf.md. Output: paths under ~/.cache/huggingface/hub/, suitable as input for mi import <DIR>.

Out-of-scope today

• Single-command HF → .model. mi download puts files in the HF cache; mi import then converts them. Composing both into one is a planned mi fetch subcommand, not yet shipped.
• Multi-shard GGUF input (*.gguf.00001-of-N). import.md declares the contract but the CLI only reads a single GGUF.
• Non-GGUF source formats at import time. The loader module supports ONNX and safetensors, but run_import only locates *.gguf files in the source directory.

These gaps are tracked; the CLI surface above is what works today.