fix: Make judge runners non-multi-turn

[jsonbailey]

Summary

Restores judge correctness lost in #166, which moved conversation history into the provider model runners.

A Judge keeps one runner across successive evaluate(...) calls. After #166, every successful run() mutated the runner's internal _history / _chat_history, so each evaluation observed the previous one's prompt and response in its history, and concurrent evaluations raced on that mutable state.

Fix

• Add a multi_turn: bool = True parameter on OpenAIModelRunner.__init__ and LangChainModelRunner.__init__. The success-path history mutation is now guarded by self._multi_turn. Default is True, so existing chat behavior is preserved.
• Thread multi_turn through RunnerFactory.create_model and through both provider create_model implementations.
• In LDAIClient._create_judge_instance, construct the judge's runner with multi_turn=False so judges share a stateless runner.
• Fix Judge.evaluate_messages to render the conversation as "{role}: {content}" on newlines (was '\r\n'.join([msg.content ...])), preserving who said what so the judge can distinguish user vs. assistant turns.

Because the new parameter is additive (default True), no caller's behavior changes outside the judge path.

Tests

• OpenAI runner: multi_turn=False keeps _history length pinned across two run() calls; default behavior still accumulates (regression guard for feat: Support conversation history directly in AI Provider model runners #166).
• LangChain runner: same shape against _chat_history.messages.
• RunnerFactory.create_model forwards multi_turn to the provider and the constructed runner reports _multi_turn == False.
• Judge integration test: a fake runner snapshots its history at call time; two evaluate() calls see the same baseline.
• evaluate_messages test: [user/hi, assistant/hello] is forwarded as "user: hi\nassistant: hello".

make test and make lint pass clean.

Refs: AIC-2544
Relates: https://github.com/launchdarkly/sdk-specs/pull/166

Test plan

• make test
• make lint
• e2e: run the chat-judge example against this branch and verify two successive judge evaluations do not see each other's history

---

Note

Medium Risk
Changes core runner construction and statefulness semantics by threading a new multi_turn flag through factories and runners; incorrect wiring could alter conversation behavior or judge correctness under concurrency.

Overview
Makes model runners optionally stateless by adding a multi_turn flag (default True) to the OpenAI and LangChain model runners and their create_model factory methods, and only appending to in-memory history when multi_turn is enabled.

Threads multi_turn through AIProvider.create_model and RunnerFactory.create_model, and updates judge creation (LDAIClient._create_judge_instance) to construct runners with multi_turn=False so repeated judge evaluations don’t contaminate each other’s prompts.

Updates Judge.evaluate_messages to include message roles when building the judge input ("{role}: {content}" joined by newlines), and adds targeted tests covering non-multi-turn behavior, flag forwarding, and role-preserving rendering.

^{Reviewed by Cursor Bugbot for commit 67dda58. Bugbot is set up for automated code reviews on this repo. Configure here.}