Add LLM judge to retrieval eval; switch chat to pypdf with 2000/400 chunks

- eval_retrieval.py: --judge MODEL lets an OpenAI model decide whether the
  top-5 chunks answer each question, regardless of source and page; reports
  Answer@5, answer-level MRR and agreement with the page check; verdicts are
  cached in eval/judge_cache.json
- On the curated goldset, gpt-6-luna rates pypdf and Docling equal
  (Answer@5 0.93 each, MRR 0.90 vs 0.92), so the chat stays on pypdf for now
- local_rag.py: chunks 2000/400, chunk headers off by default, OpenAI model
  gpt-6-luna; server.py loads ./vectorstore_pypdf_2000

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

local_rag.py

@@ -19,9 +19,10 @@ from langchain_text_splitters import RecursiveCharacterTextSplitter
 import docling_loader
 
 # Pipeline tuning: used when adding documents (rebuild index after changing).
-# Characters, not tokens: 1500 characters of German prose are roughly 400 tokens.
-CHUNK_SIZE = 1500
-CHUNK_OVERLAP = 300
+# Characters, not tokens: 2000 characters of German prose are roughly 550 tokens.
+# With the reranker, 2000/400 beat 1500/300 on both Kafka goldsets (eval/matrix_2000.out).
+CHUNK_SIZE = 2000
+CHUNK_OVERLAP = 400
 
 # Multilingual embedding model with a long input window (8192 tokens), so chunks
 # are never truncated. Changing it requires rebuilding the index.
@@ -34,7 +35,9 @@ PDF_LOADER = "pypdf"
 PROCESSED_DIR = "corpus_processed/data_split"
 # Prefix every chunk with author and title from the file name (Docling: plus section),
 # so questions like "according to Lehmann ..." can match. See docling_loader.chunk_header.
-CHUNK_HEADERS = True
+# Needed for Docling (which drops running heads); neutral to slightly harmful for pypdf,
+# whose pages carry the author in the running head anyway.
+CHUNK_HEADERS = False
 
 # Retrieval / query tuning
 RETRIEVAL_K = 5
@@ -75,7 +78,7 @@ class LocalRAG:
         vectorstore_path="./vectorstore",
         llm_provider="ollama",
         ollama_model="gpt-oss:20b",
-        openai_model="gpt-5-mini",
+        openai_model="gpt-6-luna",
         ollama_base_url="http://localhost:11434",
         pdf_loader=PDF_LOADER,
         processed_dir=PROCESSED_DIR,