…-1657)

Introduces `Evaluator` wrapping judges and JudgeConfiguration. The evaluator
runs all configured judges in parallel, warns+skips on missing judge keys, and
intentionally does NOT call tracker.trackJudgeResult — that responsibility
belongs in the managed layer. Attaches Evaluator to LDAICompletionConfig and
LDAIAgentConfig via createChat/createAgent. Adds Evaluator.noop() static factory.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…IC-2388)

Adds RunnerResult (provider-level result type without evaluations), ManagedResult
(managed-layer result with async evaluations promise), and LDAIMetricSummary (flat
metric summary including resumptionToken). Adds toolCalls and durationMs to
LDAIMetrics. TrackedChat.run() replaces invoke() returning ManagedResult with
LDAIMetricSummary built from tracker. Adds createModel() to LDAIClient/LDAIClientImpl
as the preferred replacement for createChat(). Updates chat-judge example.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… conversation management

- Add `Runner` and `AgentGraphRunner` interfaces in api/providers/Runner.ts.
  Runner.run takes a prompt string + optional output schema and returns a
  RunnerResult. AgentGraphRunner.run takes a string and returns an
  AgentGraphRunnerResult. Re-export both from api/providers/index.ts.
- Add the supporting `GraphMetrics` and `AgentGraphRunnerResult` types to
  api/graph/types.ts so AgentGraphRunner has its result shape on this branch.
- Rename `TrackedChat` -> `ManagedModel` (file + class). The constructor now
  takes a `Runner` instead of an `AIProvider`. The class is stateless: it
  owns no conversation history, and `run(prompt)` forwards the prompt
  directly to the runner. Drop `invoke()`, `_evaluateWithJudges`,
  `appendMessages`, `getMessages`, `getJudges`, `getProvider`, and the
  internal `messages` field.
- Update `LDAIClientImpl.createModel` to construct a `ManagedModel` with a
  `Runner`. The factory still produces a (deprecated) `AIProvider`, so a
  small `runnerFromAIProvider` adapter wraps it: it prepends the AIConfig's
  configured messages to the user prompt to preserve existing
  system-prompt behavior under the stateless contract.
- Mark `createChat` `@deprecated` (now delegates to `createModel`); keep
  `initChat` deprecated. Update the `LDAIClient` interface accordingly.
- Mark the `AIProvider` abstract class `@deprecated` in favor of `Runner`.
- Update `tracked-chat` and `chat-observability` examples to call
  `createModel` + `model.run()` instead of `createChat` + `chat.invoke()`.
- Rewrite the test suite for the stateless ManagedModel: prompt is passed
  through verbatim, no history is retained, ManagedResult is built from the
  RunnerResult plus the tracker's resumption token. Drop the old tests for
  `appendMessages`/`getMessages`/`invoke`.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>