get_paths_info performs no offline-mode check and HF_HUB_OFFLINE is baked
into hub constants at import, so the env guard cannot stop it in the parent
process; local-only GGUF reuse now skips the remote size verification
per-call, and a truncated cache fails the load into candidate failover
instead. Platform format gates apply only once the backend-reported
platform is fetched: the browser fallback can describe a different machine
(Mac browser against a remote Linux backend), which would wrongly report no
model; while unknown, candidates flow ungated and a chat-only backend
rejects ineligible ones at validation without consuming an attempt.
snapshot_size_bytes now uses the resolvers' complete-revision predicate
(config plus weights in the same revision) so a weight-only newer revision
is not sized while the older complete one loads. Cached-repo variant scans
are memoized per run, so a stalled repository times out once instead of
gating Send through a second identical bounded request.
ed52464558