feat: Evaluation Client — Lifecycle, Orchestration & Online Pipeline

Start with operations/evaluation/control_plane_client.py and models.py, then read EvaluationProcessor in on_demand_processor.py, EvaluationDataPlaneClient, and delete_online_evaluation_config() in online_processor.py. Trace the linked entry points and existing integration-test setup first. Done means all listed evaluator, online-config, session-data, orchestration, error-result, cleanup, and CI integration-test criteria are covered.

Problem

The SDK's EvaluationClient only exposes run(). On the control plane side, customers cannot programmatically create custom evaluators (LLM-as-a-judge configs), list available evaluators, update or delete evaluators, or manage online evaluation configs for continuous evaluation on live traffic — evaluator provisioning requires the console. On the data plane side, the starter toolkit's EvaluationProcessor provides significantly richer orchestration than run(): it fetches session data from CloudWatch independently, groups evaluators by level (SESSION vs TRACE), determines which spans to send based on evaluator level, and runs multiple evaluators with per-evaluator error handling. The toolkit also provides input validation, IAM role cleanup on delete, and typed config/result models.

Acceptance Criteria

• Customers can create, get, list, update, and delete custom evaluators
• Customers can create, get, list, update, and delete online evaluation configs
• Online evaluation config supports enable/disable toggling and sampling rate adjustment
• Typed result models with error introspection (has_error(), get_successful_results())
• Customers can fetch session trace data (spans + runtime logs) from CloudWatch for a given session and agent
• Customers can find the most recent session for an agent
• Multi-evaluator orchestration groups evaluators by level and selects appropriate spans per level
• Per-evaluator error handling — failures on one evaluator don't block others
• Online evaluation config deletion supports optional IAM execution role cleanup
• All functionality is verified via integration tests running in CI

Relevant Links

• EvaluationControlPlaneClient
• create_evaluator()
• create_online_evaluation_config()
• update_online_evaluation_config()
• EvaluationResult / EvaluationResults
• OnlineEvaluationConfig
• EvaluationProcessor
• evaluate_session()
• fetch_session_data()
• determine_spans_for_evaluator()
• execute_evaluators()
• EvaluationDataPlaneClient
• delete_online_evaluation_config()