Making LLM outputs auditable: the provider abstraction pattern

NumPath wraps LLM access behind a narrow Python Protocol for a clean provider abstraction.

It uses separate production and deterministic stub providers, selected by configuration and injected into the use case.

This lets insight generation run offline, be tested reliably, and stay swappable across vendors.

The design also improves auditability by tracing outputs back to database-backed evidence instead of model-only content.