Background GGUF loads pass local_files_only through the llama.cpp path:
the main quant resolves from the cache alone and a miss fails closed
instead of downloading, while the optional mmproj and MTP drafter
companions resolve cached-or-skipped, including on the crash replay. The
safetensors snapshot rewrite resolves against the live Studio-managed
hub cache rather than huggingface_hub's import-time default, which goes
stale when the cache location changes at runtime. The vision processor
fallback loads from config.path instead of config.identifier: identical
for ordinary loads, but a local-only rewrite must keep the fallback on
the local snapshot instead of refetching by repo id.
90f9e97302