feat(llm): RAG knowledge base over clinical guidelines with cited extractive answers

- docs/source/rag: PDF sources (KR PTSD 2023, KR depression 2024, meds module spec),
  pdf->txt converter (tables as markdown, page markers, hand-transcribed figures),
  word-level check against pdftotext; fixes HAM-D/MADRS fake headers and split rows
- config/rag: corpus txt + sources.yaml (documents, audiences, retrieval params, synonyms)
- app/llm/rag: Snowball stemmer, section-aware chunker with page citations,
  BM25 + synonym expansion + densest-passage reranking, extractive answer()
  with coverage gate, patient dose/prescribing filter, optional LLM rephrasing
  rejected when it adds content; CLI python -m app.llm.rag
- tests: gold retrieval set (hit@3 >= 15/18), off-topic, safety, faithfulness
- ADR-011

Refs #60

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

docs/source/rag/check_txt.py

@@ -0,0 +1,29 @@
+"""Сверка txt с PDF: какие слова из pdftotext не попали в txt (с учётом количества)."""
+import re
+import subprocess
+import sys
+from collections import Counter
+from pathlib import Path
+
+WORD = re.compile(r"[\wё]+", re.I)
+
+
+def words(text: str) -> Counter[str]:
+    text = re.sub(r"(\w)-\s*\n\s*(\w)", r"\1\2", text)  # переносы
+    return Counter(w.lower() for w in WORD.findall(text))
+
+
+args = sys.argv[1:]
+out = None
+if "--out" in args:
+    i = args.index("--out")
+    out = Path(args[i + 1])
+    del args[i : i + 2]
+for arg in args:
+    pdf = Path(arg)
+    ref = words(subprocess.run(["pdftotext", str(pdf), "-"], capture_output=True, text=True).stdout)
+    got = words(((out / pdf.name) if out else pdf).with_suffix(".txt").read_text())
+    lost = ref - got
+    total = sum(ref.values())
+    print(f"{pdf.name}: слов {total}, потеряно {sum(lost.values())} ({sum(lost.values()) / total:.2%})")
+    print("   ", lost.most_common(40))