fix(checkpoint): resolve HF repo Ids before conversion

[rbrisita]

Summary

ft checkpoint --model <hf-repo-id> produced an FTW checkpoint that was
missing its metadata (config.json, tokenizer.json,
tokenizer_config.json, etc), so the output was not self-contained and could
not be loaded with --model <ftw_dir>.

Root cause

convert_checkpoint assumes model_path is a filesystem path. It walks
it with os.walk and checks it with os.path.isfile / os.path.isdir, and
globs it for shards:

• python/freetoken/checkpoint/convert.py → _copy_metadata
  (os.path.isfile(model_path), os.walk(model_path))
• _source_fingerprint (glob.glob(os.path.join(model_path, "*.safetensors")))

All of these treat their argument as a local path. A Hugging Face repo
Id such as nvidia/Gemma-4-26B-A4B-NVFP4 is a Hub identifier, not a
filesystem path. So os.walk was looking for a directory named nvidia/ containing
Gemma-4-26B-A4B-NVFP4 relative to CWD.

The dangerous part is that this fails silently: os.walk on a
non-existent path yields nothing instead of raising. _copy_metadata
returned an empty list, copied_metadata was recorded as [], and the
conversion "succeeded" with only the .ftw shards written.

Reproduction

ft checkpoint --model nvidia/Gemma-4-26B-A4B-NVFP4 --out ./gemma-ftw
ls ./gemma-ftw
# only the .ftw shards + index; no config.json / tokenizer*.json, etc

Fix

Resolve the model path once, at the CLI boundary, before handing it to
convert_checkpoint:

model_path = download_hf_weight(ns.model)

download_hf_weight already handles the local-vs-repo distinction
elsewhere in FreeToken: a path that exists on disk is used as-is, and a
repo Id is fetched via snapshot_download and returned as a local
directory. convert_checkpoint then sees a real path in both cases and
needs no changes of its own.

Testing

Specs

 OS: Ubuntu 26.04 resolute
 Kernel: x86_64 Linux 7.0.0-30-generic
 Shell: bash 5.3.9
 Disk: 1TB
 CPU: 12th Gen Intel Core i7-12700H @ 20x 4.6GHz
 RAM: 64GB 
 GPU: NVIDIA GeForce RTX 3050 Laptop GPU
 VRAM: 4GB
 NVIDIA-SMI 595.91.07
 Driver Version: 595.91.07
 CUDA Version: 13.2
 freetoken version 0.1.3

ft checkpoint --model nvidia/Gemma-4-26B-A4B-NVFP4 --out ./out
Output now contains config.json, tokenizer.json,
tokenizer_config.json, etc, and
index["copied_metadata"] is non-empty.

ft checkpoint --model /local/hf-dir --out ./out
Unchanged; local paths still pass through untouched.

ft checkpoint --model /local/ftw-dir --out ./out
Unchanged; responds with "<model> is already an FTW checkpoint".

ft serve --model ./out loads the FTW via the fast path without
referencing the original HF directory.

Issue

FIxes issue #147