Hugging Face hub cache layout and shared-blob deletion hazard for older clients

Finding: The Hugging Face hub cache is a file-based store of model and dataset downloads, and parts of it are shared across repositories. The manage-cache guide says the default CACHE_DIR is ~/.cache/huggingface/hub and can be moved with HF_HOME or HF_HUB_CACHE. Within each repository folder the cache has blobs holding downloaded files named by hash, snapshots holding per-revision symlinks, refs, and a trees folder with per-commit file lists. Creation: these are written by snapshot_download and hf_hub_download when models or datasets are fetched. Growth: the guide says old revisions are kept so they can be reused, so files from earlier revisions remain until deleted; total size is UNKNOWN because no size was supplied. Measurement: the guide refers to cache inspection tools, but the supplied excerpts do not give a command, so the measurement command is UNKNOWN; inspect with the hf cache tooling described by the vendor. Safe cleanup conditions: the guide says the cache holds re-downloadable data and that huggingface_hub writes CACHEDIR.TAG so backup tools may exclude it. Cleanup should use current huggingface_hub commands, because for Xet files older clients may delete a shared blob still used by other repos, and affected files are re-downloaded next use. Lost state: deleted model files are not kept, and offline loads can fail if the local snapshot is incomplete. Regeneration: re-download from the Hub, which can be expensive for large models. Prevention: pin revisions with resolve_revision for reproducibility, and use an up-to-date client. Platform and version: symlink behavior differs on Windows; the no-symlink mode costs disk space. Limitations: this is expected model caching, not a bug; the excerpts give no size or growth measurements.