fix(gateway): snappier LLM keep-warm + higher timeout

- keep-warm interval 600s -> 240s (avoids ~47s cold-start fallback)
- local backend timeout 60s -> 70s (tolerates cold reload under 90s)

tools/zacus-gateway/main.py

@@ -241,8 +241,10 @@ async def _chat_reply(http: httpx.AsyncClient, persona: str, history: list[dict]
     """
     messages = [{"role": "system", "content": persona + _PHONE_STYLE}] + history
     backends = [
-        # 60 s tolerates one cold model (re)load; a keep-warm task keeps it ~2 s.
-        ("local", settings.local_chat_url, settings.local_chat_model, 60.0),
+        # 70 s tolerates a ~47 s cold (re)load with margin; the keep-warm task
+        # keeps it ~2 s in normal operation. Total turn (STT ~10 s + this + say
+        # ~2 s) stays under the firmware's 90 s turn_client timeout.
+        ("local", settings.local_chat_url, settings.local_chat_model, 70.0),
         ("ailiance", settings.ailiance_tts_url, settings.ailiance_chat_model, 12.0),
     ]
     last_exc: Exception | None = None
@@ -285,8 +287,12 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
 
     async def _keep_llm_warm() -> None:
         """Ping the local LLM periodically so it never auto-unloads/cools between
-        calls — the first call after idle otherwise pays a ~40 s model reload.
-        A tiny 1-token request loads (or keeps) the model resident."""
+        calls — a cold (re)load takes ~47 s, long enough to blow the per-call
+        timeout and fall back. A tiny request keeps the model resident. Ping
+        every 4 min: well under the server's 1800 s idle-unload, with margin for
+        faster cooling under memory contention. On the very first ping (gateway
+        just started) the model loads cold — that's expected and only delays the
+        first call, which the 'un instant' filler covers."""
         while True:
             try:
                 await _chat_one(
@@ -296,7 +302,7 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
                 logging.info("LLM keep-warm: %s resident", settings.local_chat_model)
             except Exception as exc:  # noqa: BLE001 — best-effort
                 logging.warning("LLM keep-warm failed: %s", type(exc).__name__)
-            await asyncio.sleep(600)  # 10 min < auto-unload-idle (1800 s)
+            await asyncio.sleep(240)  # 4 min ≪ 1800 s idle-unload, margin for contention
 
     prewarm_task = asyncio.create_task(_prewarm_greetings())
     warm_task = asyncio.create_task(_keep_llm_warm())