A forced guard now only no-ops when HF_HUB_OFFLINE itself is truthy:
huggingface_hub ignores TRANSFORMERS_OFFLINE, so with only that flag set the
GGUF cache-size check could still call get_paths_info. Snapshot selection
prefers revisions holding the inventoried safetensors weights, so a newest
metadata-only revision no longer shadows a complete older one and fails a
valid candidate. The orchestrator's parent-side preflight (transformers tier
probe and GPU sizing, which calls hf model_info) runs under the local-only
guard, closed before the worker spawn so the child does not inherit the
offline env for its whole lifetime; ordinary loads keep their current
unguarded behavior.
224b90eb02