refactor: use common LDValueConverter and LDContextEncoder in AI SDK

[mattrmc1]

Summary

Wires configured judges into completion and agent configs so online evaluation actually runs. Previously completionConfig/agentConfig always attached Evaluator.noop() regardless of a config's judgeConfiguration, so judges were dormant (intentionally descoped in v1.0 — see #180). This completes that wiring.

Applications supply a Runner per judge via a new AIRunnerProvider; the SDK builds a real Evaluator from the judgeConfiguration. Behavior is unchanged when no provider is configured — everything falls back to Evaluator.noop().

New type

@FunctionalInterface
public interface AIRunnerProvider {
  // Returns a Runner for the given judge AI Config, or null to skip this judge.
  Runner create(AIJudgeConfig judgeConfig);
}

The application implements this to wrap any model provider. The client calls it once per judge key when building an Evaluator for a completion or agent config that carries a judgeConfiguration.

New constructor

// Existing one- and two-arg constructors delegate with a null provider (online eval disabled).
public LDAIClientImpl(LDClientInterface client, LDLogger logger, AIRunnerProvider runnerProvider);

LDAIClientImpl changes

The hard-coded Evaluator.noop() in the completion and agent paths (both the resolved-variation and default paths) is replaced with buildEvaluator(judgeConfiguration, context, variables). Judge configs themselves still wire Evaluator.noop() internally — judges do not evaluate themselves.

private Evaluator buildEvaluator(JudgeConfiguration judgeConfig, LDContext context, Map<String, Object> variables) {
  if (runnerProvider == null || judgeConfig == null || judgeConfig.getJudges().isEmpty()) {
    return Evaluator.noop();
  }
  // For each judge key: fetch its AI Config (Mode.JUDGE), get a Runner from the
  // provider, build a Judge. Skip (and log) disabled judges, null runners, or any
  // construction failure so one bad judge never blocks the parent config.
  // ...
  return judges.isEmpty() ? Evaluator.noop() : new Evaluator(judges, judgeConfig, logger);
}

Per-judge sampling rates continue to flow from the JudgeConfiguration through Evaluator.evaluate — the JudgeConfiguration remains the source of truth. Judge configs are fetched through the internal evaluation path, so building an evaluator does not emit judge usage-metric events.

Fallback / isolation behavior

• Returns Evaluator.noop() when no provider is configured, the judgeConfiguration is absent/empty, or every judge fails to construct.
• A disabled judge config, a null runner from the provider, or an exception during construction skips only that judge (logged); the parent config is still built with the surviving judges.

Migration

None required. Additive only — a new interface plus a new constructor overload. Existing callers behave identically (noop evaluator). Online evaluation is opt-in by passing an AIRunnerProvider.

Test plan

• ./gradlew :lib:sdk:server-ai:test passes
• completionConfigWithJudgeConfigAndProviderBuildsRealEvaluator — real (non-noop) evaluator built
• agentConfigWithJudgeConfigAndProviderBuildsRealEvaluator — same for agent configs
• completionConfigWithNoJudgesYieldsNoopEvaluator — no judgeConfiguration → noop
• completionConfigWithNullRunnerProviderYieldsNoopEvaluator — two-arg constructor → noop even with a judgeConfiguration
• disabledJudgeConfigIsFilteredAndEvaluatorIsNoop — disabled judge skipped and logged
• nullRunnerFromProviderSkipsThatJudge — provider returns null → judge skipped, logged
• throwingRunnerProviderSkipsThatJudgeButKeepsOthers — one bad judge skipped, others survive

Additional context

• Completes the online-evaluation wiring deferred from v1.0 in feat: add Runner, RunnerResult, Judge, and Evaluator #180.
• When a provider is configured, each judge in a config's judgeConfiguration is fetched and constructed eagerly at config-retrieval time (one flag evaluation per judge key), even if the evaluator is never invoked.
• Known follow-up (out of scope): unlike js-core, Java does not yet reserve message_history / response_to_evaluate variables or strip legacy judge template messages before evaluation. Not a regression.

---

Note

Low Risk
Internal refactor behind existing parsers/interpolation; risk is mainly behavioral parity if shared encoder/converter semantics differ slightly from the removed local code.

Overview
Replaces duplicated server-ai internals with launchdarkly-java-sdk-common (2.5.0).

server-ai now depends on launchdarkly-java-sdk-common and drops the local LDValueConverter implementation (and its unit tests). AIConfigParser still maps model/tool JSON via LDValueConverter.toMap, but that type now comes from com.launchdarkly.sdk.

Interpolator no longer builds the ldctx Mustache variable with private contextToMap / singleContextToMap helpers; it uses LDContextEncoder.encode(context) instead, while keeping the same rule that caller-supplied ldctx is overridden.

No public API changes; this is internal parsing/interpolation alignment with other Java SDKs.

^{Reviewed by Cursor Bugbot for commit 3ca7b7f. Bugbot is set up for automated code reviews on this repo. Configure here.}