[YoungjaeDev]

Summary

Since #619 reuses the RAG object across requests, rag.memory carries chat turns from one request into the next.

 research_chat(request)
   rag = prepare_repo_index(request)   # cached and shared per repo/token/provider/model/filters
-  rag.memory.add_dialog_turn(...)     # accumulates on the shared object
-  history = format(rag.memory())      # includes other requests' turns, and repeats
+  history = format(request.messages)  # this request's turns only
   prompt = prompt_builder(history, context, query)

Memory and CustomConversation had no other users, so they are removed. The <turn> format and the user/assistant pairing rule are unchanged.

Evidence

tests/backend/services/test_research_chat.py runs the real RAG class and stubs only the embedder, retriever, token count, model config, and streamer.

send(question A, answer A, follow-up A)
send(question B)           -> prompt has no turns from A
send(same conversation)x2  -> each turn appears once
send(Q1, A1, Q2, A2, Q3)   -> Q1 and Q2 in order, Q3 only in <query>

• Before:

  FAILED test_request_without_history_ignores_earlier_request
  FAILED test_resending_a_conversation_does_not_duplicate_turns
  2 failed, 1 passed
  

  The third test guards turn order, so it passes both before and after.

  After:

  3 passed
  python -m pytest tests/backend tests/test_config.py tests/test_extract_repo_name.py -> 72 passed
  

The model calls are stubbed. This shows what goes into the prompt, not whether a model would have repeated a leaked conversation in its answer.

Merge Danger

Door: two-way

Blast Radius: chat

Wiki generation also calls research_chat, but it sends a single message, so its history is empty. No change to _RAG_CACHE or the Deep Research flow.