fix: Make judge runners non-multi-turn

[jsonbailey]

AIC-2544

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

Problem

After #1371 moved conversation history into the provider model runners,
every successful run() call mutates the runner's internal _history /
_chatHistory. This breaks judges: a Judge shares one runner across
successive evaluate(...) calls, so each evaluation after the first
sees the previous evaluation's prompt and response in its history.
Concurrent evaluations on the same judge also race on the mutable state.

Fix

Add a multiTurn: boolean = true parameter on every provider model
runner constructor (OpenAIModelRunner, LangChainModelRunner,
VercelModelRunner) and thread it through each provider's
*RunnerFactory.createModel and the central RunnerFactory.createModel.

When multiTurn is true (the default), behavior is unchanged — the
runner persists the user prompt and the assistant reply to its history
on successful calls. When multiTurn is false, history is treated as
read-only: every run() starts from the seeded config messages plus the
current input.

LDAIClientImpl._createJudgeInstance constructs its runner with
multiTurn=false. createModel (chat) keeps the default.

Because multiTurn defaults to true, this is additive (not breaking).
It restores the pre-#1371 contract for judges and the new parameter is
opt-in for any other stateless use case.

Also in this PR

Judge.evaluateMessages now renders each message as <role>: <content>
joined by newlines instead of dropping the role and joining contents
with \r\n. The judge model now sees who said what in the message
history section.

Test plan

• OpenAI / LangChain / Vercel runner tests cover multiTurn=false
  (no history accumulation, internal history length stays at the
  seeded length) and confirm the default still appends.
• RunnerFactory.test.ts verifies multiTurn (defaulting to true)
  is forwarded to the provider factory, and that multiTurn=false
  passes through.
• LDAIClientImpl.test.ts confirms createJudge invokes
  RunnerFactory.createModel with multiTurn=false.
• Judge.test.ts adds coverage for the role-prefixed
  evaluateMessages format and verifies successive evaluate(...)
  calls receive clean, independent input strings.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches runner construction and RunnerFactory.createModel signatures across all server AI providers; while defaults preserve existing chat behavior, incorrect threading of the new flag could change history semantics for callers (notably judges). Also changes judge prompt formatting, which can affect evaluation outputs.

Overview
Adds a multiTurn (default true) option to OpenAIModelRunner, LangChainModelRunner, and VercelModelRunner, and threads it through each provider *RunnerFactory.createModel, the base AIProvider.createModel, and the central RunnerFactory.createModel; when false, successful run() calls no longer persist user/assistant messages to internal history.

Updates judge initialization (LDAIClientImpl.createJudge/_createJudgeInstance) to create runners with multiTurn=false, preventing cross-evaluation history bleed/races. Judge.evaluateMessages now formats history as "<role>: <content>" joined by \n instead of dropping roles.

Extends unit tests across providers and SDK (RunnerFactory, LDAIClientImpl, Judge) to cover stateless behavior and default multiTurn=true forwarding.

^{Reviewed by Cursor Bugbot for commit ce73d8c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

@launchdarkly/js-sdk-common size report
This is the brotli compressed size of the ESM build.
Compressed size: 26208 bytes
Compressed size limit: 29000
Uncompressed size: 128789 bytes

[github-actions]

@launchdarkly/js-client-sdk size report
This is the brotli compressed size of the ESM build.
Compressed size: 31906 bytes
Compressed size limit: 34000
Uncompressed size: 113658 bytes

[github-actions]

@launchdarkly/browser size report
This is the brotli compressed size of the ESM build.
Compressed size: 179498 bytes
Compressed size limit: 200000
Uncompressed size: 830837 bytes

[github-actions]

@launchdarkly/js-client-sdk-common size report
This is the brotli compressed size of the ESM build.
Compressed size: 38487 bytes
Compressed size limit: 39000
Uncompressed size: 211236 bytes