feat: port template upgrades (citations, eval, chunking, nomic prefixes)

mcp 2.x already on origin/main (#10). Does not change BM25-first
search_docs — n=6 is too small to flip the default.

- Numbered [1] citations via docs_mcp/format.py
- Eval P@1 + JSONL sidecar + eval.pvalue + eval.trace
- Heading-recursive chunker, keep chunk-0 and MAX_CHARS=4000
- Nomic prefixes at embed time only; stored text unprefixed

Closes #12

eval/retrievers.py

@@ -51,7 +51,9 @@ class DenseRetriever:
         self.pool = pool
 
     def retrieve(self, query: str, k: int = 10) -> list[tuple[str, str]]:
-        res = self.col.query(query_texts=[query], n_results=self.pool)
+        from rag.embeddings import EMBED_QUERY_PREFIX, embed_texts
+        qvec = embed_texts([query], prefix=EMBED_QUERY_PREFIX)[0]
+        res = self.col.query(query_embeddings=[qvec], n_results=self.pool)
         ids = (res.get("ids") or [[]])[0]
         return _collapse_to_pages(ids, k)