fix(gateway): re-baseline agent-cache message_count after each turn

The #45966 cross-process coherence guard snapshots a session's on-disk
message_count next to the cached agent and rebuilds the agent when the
count changes.  But the snapshot is taken at agent-BUILD time — before
the turn writes its own user + assistant (+ tool) rows — and the cache
entry is never rewritten on a reuse.  So this process's OWN turn grows
message_count, and the very next turn sees a mismatch and rebuilds the
agent.  That happens every turn, for every conversation, silently
destroying the per-conversation prompt caching the cache exists to
protect (AGENTS.md: prompt caching is sacred).

Add _refresh_agent_cache_message_count(): after a turn completes and the
agent has flushed its rows to the SessionDB, re-baseline the stored count
to the now-current value.  The guard then fires ONLY when a DIFFERENT
process changes the transcript — preserving the #45966 fix while keeping
the cache warm for normal single-process operation.

Tests drive the real SessionDB + the real guard condition: 5 consecutive
same-process turns now all REUSE the cached agent (0 before the fix); a
cross-process append still invalidates; and the re-baseline is fail-safe
(no DB, falsy session_id, raising probe, legacy 2-tuple, pending sentinel
all no-op).

gateway/run.py

@@ -8875,6 +8875,20 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                 _response_time, _api_calls, _resp_len,
             )
 
+            # Re-baseline the cached agent's message_count snapshot now that
+            # this turn has completed and the agent has flushed its rows to
+            # the SessionDB.  The cross-process coherence guard (#45966)
+            # snapshots the count at agent-BUILD time (before this turn's own
+            # writes) and never refreshes it on reuse — so without this, this
+            # process's own turn would grow the count and the next turn would
+            # see a mismatch and rebuild the agent every turn, destroying
+            # prompt caching.  Refreshing here makes the guard fire only on a
+            # DIFFERENT process's writes.  Uses the (possibly compaction-
+            # updated) live session_id.  Fail-safe inside the helper.
+            self._refresh_agent_cache_message_count(
+                session_key, session_entry.session_id
+            )
+
             # Successful turn — clear any stuck-loop counter for this session.
             # This ensures the counter only accumulates across CONSECUTIVE
             # restarts where the session was active (never completed).
@@ -12784,6 +12798,57 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
         if release_running_state:
             self._release_running_agent_state(session_key)
 
+    def _refresh_agent_cache_message_count(
+        self, session_key: str, session_id: Optional[str]
+    ) -> None:
+        """Re-baseline a cached agent's stored message_count after THIS turn.
+
+        The cross-process coherence guard (#45966) compares the session's
+        on-disk ``message_count`` against the count snapshotted next to the
+        cached agent, and rebuilds the agent on a mismatch.  But the snapshot
+        is taken at agent-BUILD time — before this turn writes its own user +
+        assistant (+ tool) rows — and the cache entry is never rewritten on a
+        reuse.  So without this re-baseline, THIS process's own turn would
+        grow ``message_count`` and the very next turn would see a mismatch
+        and rebuild the agent — every turn, for every conversation — silently
+        destroying the per-conversation prompt caching the cache exists to
+        protect.
+
+        Call this once a turn has completed and the agent has flushed its
+        rows to the SessionDB.  It snapshots the now-current count (which
+        includes this process's own writes) so the guard only fires when a
+        DIFFERENT process changes the transcript out from under us.  The
+        ``_sig`` is left untouched; only the count element is refreshed, and
+        only when the same agent is still cached (no rebuild/eviction raced
+        in between).  Fail-safe: any DB error leaves the snapshot as-is, which
+        at worst costs one unnecessary rebuild on the next turn.
+        """
+        if self._session_db is None or not session_id:
+            return
+        _cache_lock = getattr(self, "_agent_cache_lock", None)
+        _cache = getattr(self, "_agent_cache", None)
+        if not _cache_lock or _cache is None:
+            return
+        try:
+            _sess_row = self._session_db.get_session(session_id)
+            _live = _sess_row.get("message_count", 0) if _sess_row else None
+        except Exception:
+            return
+        if _live is None:
+            return
+        with _cache_lock:
+            cached = _cache.get(session_key)
+            # Only re-baseline a live 3-tuple entry; skip pending sentinels,
+            # legacy 2-tuples (they intentionally opt out of the guard), and
+            # the case where the entry was evicted/rebuilt mid-turn.
+            if (
+                isinstance(cached, tuple)
+                and len(cached) > 2
+                and cached[0] is not _AGENT_PENDING_SENTINEL
+            ):
+                if cached[2] != _live:
+                    _cache[session_key] = (cached[0], cached[1], _live)
+
     def _evict_cached_agent(self, session_key: str) -> None:
         """Remove a cached agent for a session (called on /new, /model, etc).