docs: alinear modelos a qwen2.5:7b + bge-m3 (iGPU activa)

Generación 3B→7B y embeddings a bge-m3 en README, CLAUDE.md, .env.example y
k8s/deployment.yaml. Corregido el default buggy de config.py: ollama_embed_model
era qwen2.5:3b (un modelo de chat) → bge-m3. Añadido OLLAMA_EMBED_MODEL donde
faltaba. Nota del host público de ollama ahora tras Authentik.

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

README.md

@@ -18,7 +18,7 @@ ExhaustiveScraper
     ├── PDFs (public documents)
     └── Web scraping (trafilatura)
          ↓ recursive expansion (depth 1-3)
-ContentProcessor (Ollama qwen2.5:3b)
+ContentProcessor (Ollama qwen2.5:7b + bge-m3 embeddings)
     ├── Chunking (800 token chunks, 100 overlap)
     ├── Quality scoring (0-10 per chunk)
     ├── Embeddings (cosine similarity RAG)
@@ -102,7 +102,7 @@ git add . && git commit -m "feat: add researchowl" && git push
 
 ## Notes
 
-- Uses **qwen2.5:3b** (your existing Ollama) for all AI tasks — zero API cost
+- Uses **qwen2.5:7b** (scoring) and **bge-m3** (embeddings) on your existing Ollama — zero API cost
 - Optionally add `ANTHROPIC_API_KEY` for Claude fallback on generation
 - SQLite database stored in `/data/researchowl.db`
 - All outputs saved to DB and available via `/outputs`