Defines an abstract hierarchy (`Metric`, `MetricInput`, `MetricComputation`) for computing monitoring metrics such as feature drift, output drift, data quality, and runtime health. The base classes encapsulate the compute-then-threshold pattern with a registry for discoverability. This pattern is reusable in any monitoring system that needs to add custom metrics without modifying the framework.

Implements `ForkNodeHandler`, which reads experiment proposals from multiple file sources (`.proposals/`, `.luml-fork.json`, `.prisma/fork.json`) to automatically branch and spawn child experiment runs. The handler cleanly separates proposal parsing, validation, and node creation, demonstrating a robust design for agent-driven iteration that could be extracted for any automated A/B experimentation system.

Provides `LLMJudgeScorer` and `SupervisedLLMJudgeScorer` abstract classes that use a configurable LLM to evaluate model outputs. Includes OpenTelemetry tracing, automatic retry on parse failures with corrective reminders, and a name disambiguation mechanism for multiple scorers. This base is cleanly separated from concrete scorers (correctness, relevancy, etc.) and can be reused to define custom LLM-judge scorers in any project.

Implements a generic request-response proxy over a Web Worker using unique message IDs, supporting `AbortSignal` cancellation and a lifecycle `init` method. It provides a typed interface for the frontend to communicate with an experiment snapshot worker without exposing worker internals. This pattern is directly reusable for any single-purpose Web Worker communication (e.g., in-browser ML, data processing) and demonstrates clean separation of concerns between UI and worker logic.