fix: json.dumps dict before psycopg2 insert + remove fatal raise in save_context

save_context/_save_action_plan passed raw Python dicts as SQL bind params,
causing psycopg2.ProgrammingError that propagated via raise and crashed the
entire AI pipeline, forcing every natural language message to keyword fallback.

Also increase Hermes intent timeout 15s→30s for qwen2.5 cold-start latency.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

services/ai_orchestrator.py

@@ -1,5 +1,6 @@
 # services/ai_orchestrator.py
 import asyncio
+import json
 import logging
 from typing import Any, Dict, Optional
 
@@ -73,14 +74,13 @@ class AIOrchestrator:
                     "sid": session_id,
                     "ag": agent,
                     "ck": "latest",
-                    "cv": payload,
+                    "cv": json.dumps(payload, ensure_ascii=False),
                 },
             )
             session.commit()
         except Exception as e:
             session.rollback()
-            logger.error(f"[AIOrchestrator] save_context failed: {e}")
-            raise
+            logger.warning(f"[AIOrchestrator] save_context failed (non-fatal): {e}")
         finally:
             session.close()
 
@@ -98,13 +98,12 @@ class AIOrchestrator:
                     "sid": plan.get("session_id"),
                     "pt": plan.get("plan_type"),
                     "sku": plan.get("sku"),
-                    "pl": plan,
+                    "pl": json.dumps(plan, ensure_ascii=False),
                 },
             )
             session.commit()
         except Exception as e:
             session.rollback()
-            logger.error(f"[AIOrchestrator] save_action_plan failed: {e}")
-            raise
+            logger.warning(f"[AIOrchestrator] save_action_plan failed (non-fatal): {e}")
         finally:
             session.close()

services/hermes_analyst_service.py

@@ -154,7 +154,7 @@ class HermesAnalystService:
             resp = requests.post(
                 f"{HERMES_URL}/api/generate",
                 json=payload,
-                timeout=15,  # 意圖分類不必等 120s
+                timeout=30,  # 意圖分類，qwen2.5 首次推理可能需 ~20s
             )
             resp.raise_for_status()
             raw = (resp.json().get("response", "") or "").strip()