feat: Hugging Face model browser — search, download, and manage models from HF

[rexlunae]

Inspiration: Unsloth

Unsloth lets you search + download + run any model (GGUF, safetensors, LoRA adapters) from Hugging Face directly within the UI. It also auto-detects pre-existing model downloads from HF cache, LM Studio, etc.

What RustyClaw could do

Add a Hugging Face integration layer:

1. Search models: rustyclaw model search "qwen 7b" — searches HF with filters (task, size, license, likes)
2. Download models: rustyclaw model pull unsloth/Qwen3.5-4B-GGUF:UD-Q4_K_XL — downloads to local cache
3. Auto-detect existing models: Scan ~/.cache/huggingface/, ~/.cache/lm-studio/, Ollama's model dir for already-downloaded models
4. Model info: rustyclaw model info <id> — shows size, quant, context length, recommended params, license
5. List local models: rustyclaw model list — all locally available models across Ollama, HF cache, etc.

Integration with existing systems

RustyClaw already has:

• ollama_manage tool for Ollama model lifecycle
• Provider registry with Ollama as a provider
• Model registry in models/registry.rs

This would add a Hugging Face client that complements (not replaces) the Ollama path. A model downloaded from HF could be served via Ollama or llama.cpp.

Scope

• rustyclaw-core: HuggingFace client (using hf-hub crate or raw API)
• rustyclaw-cli: rustyclaw model subcommand with search/pull/list/info
• Caching layer respecting HF_HOME/HF_HUB_CACHE env vars
• Optional hf feature flag

Prior art

• Unsloth Studio's model search + download UX
• huggingface_hub Python library
• hf-hub Rust crate: https://crates.io/crates/hf-hub
• Ollama's model pull functionality