fix: correcciones de scraping/DB y mejoras de robustez

Sección crítica:
- is_blacklisted: match por dominio/subdominio exacto (antes "x.com" como
  substring bloqueaba netflix.com, phoenix.com, etc.)
- normalize_url: conserva el query string (rompía YouTube watch?v= y URLs
  con ?id=); solo borra el fragment
- get_db: PRAGMA busy_timeout=5000 para evitar "database is locked" en
  /compare y watches solapados
- OllamaClient.embed: usa OLLAMA_EMBED_MODEL en vez del modelo de chat
- log_api_call: coste por modelo (opus/sonnet/haiku) en vez de Haiku fijo

Mejoras:
- src/llm.py: cliente Anthropic compartido y cacheado (antes se instanciaba
  uno por cada llamada/chunk)
- SEARXNG_URL configurable via env
- get_running_loop() en vez de get_event_loop() (deprecado)
- soup.title.get_text() robusto ante <title> con tags anidados
- limpieza: import muerto, total_words duplicado, w_id no usado

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

src/processor/processor.py

@@ -23,6 +23,7 @@ class OllamaClient:
     def __init__(self):
         self.base_url = settings.ollama_url.rstrip("/")
         self.model = settings.ollama_model
+        self.embed_model = settings.ollama_embed_model
 
     async def generate(self, prompt: str, system: str = None,
                        timeout: int = 120, temperature: float = 0.7) -> str:
@@ -47,7 +48,7 @@ class OllamaClient:
 
     async def embed(self, text: str) -> Optional[list[float]]:
         """Get embedding vector for a text"""
-        payload = {"model": self.model, "prompt": text}
+        payload = {"model": self.embed_model, "prompt": text}
         try:
             async with httpx.AsyncClient(timeout=60) as client:
                 resp = await client.post(f"{self.base_url}/api/embeddings", json=payload)
@@ -124,7 +125,6 @@ class ContentProcessor:
     async def process_session(self, session_id: int, topic: str,
                                progress_callback=None) -> dict:
         """Process all scraped sources for a session"""
-        from src.db.database import ResearchDB
         sources = await self.db.get_all_sources(session_id)
         scraped = [s for s in sources if s["status"] == "scraped"]
 
@@ -132,7 +132,6 @@ class ContentProcessor:
         scraped = await self._dedup_sources(session_id, scraped)
         logger.info("After dedup", unique=len(scraped))
         total_chunks = 0
-        total_words = 0
 
         semaphore = asyncio.Semaphore(3)  # process 3 sources at once
 
@@ -276,7 +275,7 @@ class ContentProcessor:
 
     async def _score_with_claude(self, chunk: str, topic: str,
                                   session_id: int | None = None) -> float:
-        import anthropic
+        from src.llm import get_anthropic_client
         prompt = (
             f'Rate 0-10 how relevant this text is to the topic "{topic}". '
             f'Be generous — if the text is tangentially related, score 4+. '
@@ -284,7 +283,7 @@ class ContentProcessor:
             f'Reply with only a number.\n\nText:\n{chunk[:500]}'
         )
         try:
-            client = anthropic.AsyncAnthropic(api_key=settings.anthropic_api_key)
+            client = get_anthropic_client()
             msg = await client.messages.create(
                 model=settings.claude_model,
                 max_tokens=10,