docs(openspec): propose creating an Outfit from a Hugging Face ref

[outofcoffee]

An OpenSpec change proposing outfit hf <ref>, a command that turns a Hugging Face model reference into a working Outfit.

Summary

• New command outfit hf <ref>: reads a Hugging Face model repo and writes the Outfit for it, so a pasted model reference becomes a working config without hunting through the model page for the quantisation, the engine and the context window.
• Everything is inferred and every inference is overridable: PROVIDER from the repo's files (GGUF → llamacpp, MLX → omlx, safetensors → vllm), the quantisation from the GGUF filenames, CONTEXT from the published config.json, and ALIAS from the repo name. The reasoning goes to stderr so stdout stays a clean Outfit.
• The existing caches are read, never bypassed: a model already in the Hugging Face cache — or in llama.cpp's own cache — is named by its path on disk, so the engine loads what is there instead of downloading a second copy. A fully cached model resolves with no network request at all.
• outfit never downloads weights. Describing a model is metadata only; fetching stays the engine's job, as it is today.
• Optional Hugging Face token resolved from HF_TOKEN, HUGGING_FACE_HUB_TOKEN or the CLI's token file, so gated and private repos work for a logged-in user. It is never written or printed.
• --output-file/-o writes the file (refusing to clobber without --force), --apply dresses the harness in the same step, and --no-cache forces the portable repo-reference form.
• Two new capabilities: huggingface-hub (reference forms, Hub reads, token, cache lookup) and huggingface-outfits (the command and its inference rules).

Implementation details

Two decisions are worth a reviewer's attention.

No Go SDK. The closest thing is gomlx/go-huggingface, which does share the Python cache layout, but its module requires gomlx, parquet-go, protobuf, sentencepiece and lipgloss — a machine-learning framework's dependency graph for a CLI with three direct dependencies. The other candidates are downloader-shaped, and the download path is exactly what this change does not want. internal/hf is stdlib-only: two JSON endpoints and a documented directory layout. The package boundary is where a library could slot in later if the write path is ever wanted.

The two caches are separate. llama-server -hf writes into LLAMA_CACHE, not the Hugging Face cache, so a model downloaded by outfit serve is invisible to the HF layout and vice versa — both are checked. llama.cpp's filename convention is not a documented contract, so that lookup requires the owner, the repo name and the quant to all match; a false negative just falls back to the repo reference, which is what would have been written anyway.

One consequence worth naming: a cached-path MODEL describes one machine's disk, and Outfits do get committed. The narration says so at the moment it happens, and --no-cache produces the shareable form.