feat: Support conversation history directly in AI Provider model runners

[jsonbailey]

Summary

• OpenAI: Adds a _history: List[LDMessage] field to OpenAIModelRunner. Each successful run() appends the user message and assistant response so subsequent calls include prior turns. OpenAI Chat Completions has no built-in state management, so a manual list is the standard pattern.
• LangChain: Adds InMemoryChatMessageHistory to LangChainModelRunner. Config messages are converted to native BaseMessage types once per call and joined with the history before invoking the model. add_user_message() / add_ai_message() keep history as native LangChain types throughout, avoiding repeated conversion.
• History is only updated on successful runs — failed or empty responses leave history unchanged.
• Tests added for multi-turn accumulation, no-accumulation-on-failure, and config-message ordering.

Test plan

• OpenAI history accumulates across two successful calls
• OpenAI history not updated on exception
• LangChain history accumulates across two successful calls using InMemoryChatMessageHistory
• LangChain history not updated on exception
• Config messages still prepended before history on every call
• CI green

🤖 Generated with Claude Code

---

Note

Medium Risk
Adds stateful, cross-call message history to both model runners, which can change prompt composition and token usage for any multi-turn caller. Risk is moderate due to behavior change and potential memory growth, but updates are gated to successful, non-empty responses and covered by new tests.

Overview
Adds multi-turn conversation state to model runners. OpenAIModelRunner now keeps an internal _history (seeded with any config_messages) and appends successful user+assistant turns so subsequent run() calls include prior context.

LangChain runner now uses LangChain-native history. LangChainModelRunner replaces the one-off config_messages prepend with an InMemoryChatMessageHistory, invokes the LLM with accumulated BaseMessage history plus the new HumanMessage, and appends to history only on successful, non-empty responses.

Tests expanded to verify history accumulation, no accumulation on failures, and correct ordering of config/system messages ahead of prior turns.

^{Reviewed by Cursor Bugbot for commit 023c248. Bugbot is set up for automated code reviews on this repo. Configure here.}