fix(ingestion): resolve Slack authors by real_name, not display_name - #580

Slack messages store author ids, so every speaker, @mention, and DM label
is resolved through the export's users.json roster. That lookup preferred
profile.display_name, which in a real export is frequently a short handle
("morgan"). Zep merges entities by the names it sees in text, so a handle
never merges with the same person written in full ("Morgan Lee") in an
email or document: one person silently becomes two nodes, and half their
facts hang off each. Verified against a live graph — ingesting a
handle-based export alongside full-name documents produced both 'morgan'
and 'Morgan Lee' as Person nodes; after this change it produces one.

Prefer real_name, then display_name, then the username, then the raw id.
Slack's own precedence is the opposite, but it optimizes for how a name
reads in a chat client, not for entity resolution.

Report authors whose roster entry has no real_name, counting only those
whose content was actually ingested (including via an @mention) so a
handle-only user who never posted is not noise. Expose the raw Slack id
as SlackMessage.user_id so formatter= can substitute names from a
directory of your own when the roster is thin.

Both fixtures previously wrote display_name so it could never behave like
a display name — the example roster left it empty and the test roster put
a full name in it — which is why this went unnoticed. They now carry
realistic handles, making the fixtures a regression guard: reverting only
the loader change fails nine tests.

Co-Authored-By: Claude <noreply@anthropic.com>

… label

A DM or group DM is labeled by its members, and that label is written into
every episode's text and metadata. It is built straight from the roster in
_label(), never through _resolve(), so a member whose only name is a handle
was put into the graph untracked: if they never authored a message and were
never @mentioned, _weak_names stayed empty and no warning was emitted.

Carry the roster ids a label names on _Conversation and record the weak ones
in _load_conversation, which is the first point that knows the conversation
was both selected and non-empty. Recording them in _label() instead would
warn about members of a conversation the run skipped — a private DM excluded
by the default conversation_types — and recording them on selection alone
would warn for a selected folder that yielded no episodes.

The warning now says "named in ingested content" rather than "whose content
was ingested", which no longer fits a member who never posted.

Co-Authored-By: Claude <noreply@anthropic.com>

…kept

_normalize_text resolves @mentions, and it runs before a message is known to
be usable: the ts / thread_ts checks come after it, and the duplicate-ts drop
happens later still in _load_conversation. A handle mentioned only by a
message this run threw away therefore landed in _weak_names, and the warning
claimed it was named in ingested content when nothing of it reached the graph.

Buffer weak names per message in _parse and promote them in
_load_conversation once the message is accepted. Buffering rather than
reordering _parse keeps the existing skip precedence intact — moving the ts
checks above the empty-text check would start counting _invalid_ts for
messages dropped for having no text.

_unresolved_users deliberately keeps recording immediately: it reports ids
"referenced in messages", which is true whether or not the message survived,
whereas a weak name is reported as "named in ingested content".

Co-Authored-By: Claude <noreply@anthropic.com>