AbstractCore v2.15.1

A download that stops on its own now ends failed with a plain reason, and cancelled only after
a real cancel request, which records who made it. The fit warning states the two numbers its
verdict compared, and a copied Hugging Face cache with missing links reads absent.

Fixed

• A model's "may not fit" warning now states the two numbers its verdict compared: the total need
  (weights plus working memory and cache) against the memory a model can use (the ceiling minus the
  part kept free for the system). A 24 GiB Mac read "needs about 16 GiB; can give about 18 GiB: may
  not fit"; it now reads "needs about 16.4 GiB in total (15.2 GiB of weights plus 1.2 GiB ...); this
  computer can give a model about 16.0 GiB (... 18.0 GiB, and 2.0 GiB of that is kept free ...)".
  The fit block adds usable_bytes, reserve_bytes and overhead_bytes.
• A download that stops on its own is never reported "cancelled". Host jobs record who asked for a
  cancel (cancelled_by: console, api, cli, other_process; cancelled_by_user) and say
  why a job ended in one plain sentence (ended_reason): a dropped connection, a Hub error, a full
  disk, the owning process restarting, a failed MTP companion. A job whose owner exited reads
  failed in every field (its state said downloading).
• The Hugging Face download child stops when the process that started it exits (it kept
  downloading after a gateway restart and held the file lock a retry then waited on).
• Download messages no longer promise a resume huggingface_hub does not do: files that finished are
  kept; the file that was in progress starts over. A failed download's instruction follows from what
  failed (it always talked about gated models and an environment variable).
• A Hugging Face cache copied without its symbolic links (rsync without -a) or without its blobs is
  no longer reported installed; the model reads "absent" with a plain reason, and a download repairs
  the missing links without fetching the data again.