feat(costs): Pro-Nutzer-Kostenerfassung Stufe 1 (OpenRouter-Ist-Kosten)

CostMeter via ContextVars (app/core/costs.py, multitool-ready). Alle
OpenRouter-Calls (LLM, Sonar, Koreferenz, Fallback) senden usage:{include:true}
und buchen die Ist-Kosten je Kategorie (llm/web_search) — inkl. finalem
Streaming-usage-Chunk. cost_usage-Tabelle (user/day/category) + Store-Methoden;
ws.py/chat.py start_meter pro Turn + persist_costs. Statistik: Spalten
"Kosten (Monat)" + "Ø/Anfrage" + Gesamt. Live: 1 Such-Turn ~ $0,0052
(web_search dominiert, LLM ~ $0,00007). Tests: 318.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/api/chat.py

@@ -16,7 +16,8 @@ from app.dependencies import (
     voice_for_route,
 )
 from app.core.memory_extractor import maybe_schedule_extraction
-from app.quota import enforce_quota, record_usage, QuotaExceededError
+from app.core.costs import start_meter, persist_costs
+from app.quota import enforce_quota, record_usage, QuotaExceededError, record_web_search
 from app.runtime_config import runtime_settings
 from app.safety.emergency import (
     record_manual_emergency,
@@ -97,6 +98,7 @@ async def chat(
     # Explizit angeforderte Stimme gewinnt; sonst folgt sie der Routensprache.
     voice = payload.voice or voice_for_route(route.tts_provider, route.language, route.voice_gender)
 
+    start_meter()   # Kosten dieses Turns sammeln
     try:
         trace, audio = await orchestrator.chat_text(
             payload.text,
@@ -110,6 +112,8 @@ async def chat(
         raise HTTPException(status_code=502, detail=str(exc))
 
     record_usage(user, store, len(payload.text) + len(trace.semantic_response or ""))
+    record_web_search(user, store, trace.web_searches)
+    persist_costs(user, store)
 
     # Turn persistieren (User-Eingabe + semantische Antwort) fuer das Gedaechtnis.
     if session_id: