fix(hub): classify Hub failures by exception type, not class name

[stisiTT]

classify_hub_error compared type(exc).__name__ against "EntryNotFoundError".
huggingface_hub 1.x raises subclasses for the cases it can be specific about — a
missing file is RemoteEntryNotFoundError — so that branch silently stopped
matching and the missing-file case fell through to the repo-level 404: "no such
bundle, or no access" for a repo that plainly exists, with advice to go find a
different id or log in.

isinstance covers the subclasses. LocalEntryNotFoundError has to stay ahead
of the missing-file branch because it is also an EntryNotFoundError.

The test helper now builds the real exception class whenever the installed hub
ships one. With isinstance branching a same-named fake gets routed by the
message heuristics instead, so all of these tests would have kept passing while
the type evidence went unexercised — which is how the bug survived.

• Two new cases carry neither the manifest filename nor a network keyword, so the
  class is the only evidence (the first fails before, passes after)
• Full offline suite: 918 passed

Stacked on #130.

🤖 Generated with Claude Code

[ezietlowTT]

Approve. Reviewed and tested locally (full offline suite: 918 passed; targeted hub/publish tests: 41 passed).

The core fix is right and matters: classify_hub_error now branches with isinstance instead of matching type(exc).__name__. huggingface_hub 1.x raises RemoteEntryNotFoundError / LocalEntryNotFoundError (subclasses of EntryNotFoundError), so the old name comparison would stop matching the moment the library got more specific — silently, falling through to the repo-level 404 and telling a user their repo doesn't exist when only one file inside it is missing.

Verified against the installed hub (1.27.0):

• LocalEntryNotFoundError really is a subclass of EntryNotFoundError, so the offline branch must stay ahead of the missing-file branch — and it does. Good catch and correct ordering.
• RemoteEntryNotFoundError exists here, so the new subclass test actually runs rather than skipping.
• The imports (huggingface_hub.utils.{EntryNotFoundError, LocalEntryNotFoundError}) resolve on the >=0.23 floor.

The tests are the strong part: _exc now instantiates the real hub class when present (falling back to a fabricated one otherwise), so the isinstance path is genuinely exercised rather than passing via message heuristics; and the two new cases cover the subclass and the offline-ordering hazard with messages that carry no filename/network keyword, so the class is the only evidence.

The bundled cli.py change (publish/unpublish moving from typer.secho to console.note/console.milestone) is a reasonable consistency cleanup and has its own gutter/marker test. OfflineModeIsEnabled staying a name match is fine — it isn't reliably importable across versions.

community-labeling red is the known #80 org-secret gating, unrelated.