Sofortiger Cancel: Laufende Claude-Prozesse per Event abbrechen

Bisher war Cancel kooperativ (Flag-basiert) -- der Code pruefte das Flag
nur an wenigen Checkpoints. Laufende Claude CLI Subprozesse (WebSearch,
Analyse, Faktencheck) liefen bis zum Ende weiter, was minutenlanges
Warten beim Abbrechen verursachte.

Neuer Ansatz:
- ContextVar _cancel_event_var in claude_client.py
- Orchestrator setzt asyncio.Event vor jedem Refresh
- call_claude wartet parallel auf Prozess UND cancel_event
- Bei Cancel: process.kill() + CancelledError sofort
- Kein Durchreichen durch Agent-Methoden noetig (contextvars)

src/agents/orchestrator.py

@@ -10,7 +10,7 @@ from urllib.parse import urlparse, urlunparse, quote_plus
 
 import httpx
 
-from agents.claude_client import UsageAccumulator
+from agents.claude_client import UsageAccumulator, _cancel_event_var
 from agents.factchecker import find_matching_claim, deduplicate_new_facts, TWOPHASE_MIN_FACTS
 from source_rules import (
     _detect_category,
@@ -398,6 +398,7 @@ class AgentOrchestrator:
         self._ws_manager = None
         self._queued_ids: set[int] = set()
         self._cancel_requested: set[int] = set()
+        self._cancel_event: asyncio.Event | None = None
 
     def set_ws_manager(self, ws_manager):
         """WebSocket-Manager setzen für Echtzeit-Updates."""
@@ -441,6 +442,8 @@ class AgentOrchestrator:
         # Check if it's the currently running task
         if self._current_task == incident_id:
             self._cancel_requested.add(incident_id)
+            if self._cancel_event:
+                self._cancel_event.set()
             logger.info(f"Cancel angefordert fuer laufende Lage {incident_id}")
             if self._ws_manager:
                 try:
@@ -512,6 +515,8 @@ class AgentOrchestrator:
                 user_id = None
             self._queued_ids.discard(incident_id)
             self._current_task = incident_id
+            self._cancel_event = asyncio.Event()
+            _cancel_event_var.set(self._cancel_event)
             logger.info(f"Starte Refresh für Lage {incident_id} (Trigger: {trigger_type})")
 
             RETRY_DELAYS = [0, 120, 300]  # Sekunden: sofort, 2min, 5min
@@ -585,6 +590,8 @@ class AgentOrchestrator:
                     }, _vis, _cb, _tid)
             finally:
                 self._current_task = None
+                self._cancel_event = None
+                _cancel_event_var.set(None)
                 self._queue.task_done()
 
     async def _mark_refresh_cancelled(self, incident_id: int):