fix(server): search every settled message, not just each turn's last one

[omegent-app]

Thread search restricted assistant matches to the single row a turn names as its
terminal message:

messages.message_id IN (
  SELECT turns.assistant_message_id FROM projection_turns AS turns
  WHERE turns.assistant_message_id IS NOT NULL
)

Providers stream an answer as many assistant items — roughly ten per turn now,
and climbing as chunking gets finer — and only the last is linked as
turns.assistant_message_id. Everything said before it was invisible to search.
That is why full-text search appeared to "stop working" without anything
visibly changing: the query never changed, the shape of the data under it did.

Measured on a live state.sqlite

	searchable
before	7,118 / 20,094 settled messages (35%)
after	20,094 (100%)
assistant messages before	3,430 / 16,406 (21%)

A real query for worktree goes from 138 to 187 threads.

All 425 threads with assistant text had some searchable row, so the symptom
was a search quietly missing most of what was said, not one missing whole
threads — which is exactly why it was hard to pin down.

Scope

• Tool calls were never involved. They are activities in
  projection_thread_activities, not rows in projection_thread_messages, so
  they stay out by construction.
• system notices are now excluded explicitly rather than incidentally.
• Per-thread dedupe is unchanged: thread_match_rank = 1 still collapses a
  thread to one best row, preferring a user match, then the most recent.
• Streaming rows stay excluded via the existing is_streaming = 0.

Known limit

A phrase that straddles a split boundary still will not match — each message row
is matched independently, so "the quick brown" in one item and "fox" in the next
is two rows, not one string. Fixing that means matching against a per-turn
concatenation, which is a bigger change with its own cost; happy to do it as a
follow-up if it matters in practice.

Tests

The existing test asserted the old behaviour outright — its fixture literally
read 'Interim needle must not be searchable.' — so that expectation is
inverted here, and a split-answer case is added whose matching half is not the
turn's assistant_message_id. Mutation-checked: restoring the old restriction
fails both new assertions.

Co-authored-by: Patrick Roza 42661+patroza@users.noreply.github.com