The load route's sidecar tier probe reached _remote_lora_base's raw HTTP
request and Hub config reads, and the training guard's memory estimation
can call hf model_info; neither honors offline mode, so both now run
against the resolved snapshot path for local-only loads (a local path
resolves from config.json on disk and skips the remote adapter probe).
GGUF cached rows select a GGUF-bearing snapshot (top level or one folder
deep) instead of reusing the safetensors-preferring model helper, so a
mixed repo's safetensors revision cannot become the GGUF row's load target
while the UI reports a quant.
3026ff9e79