Add or document a reusable full-pipeline runner for mixed RAG adapters

[laynepenney]

Finding

The adapter pattern works for a mixed retrieval + generation customer flow, but there is no reusable runner that wires the two surfaces together.

Validation

A customer-shaped probe passed with three adapter shapes:

1. retrieval-only: RetrievalAdapter + run_retrieval_category
2. generation-only: GenerationAdapter + run_generation_category
3. mixed/RAG: retrieve candidates, convert them to context manually, then call GenerationAdapter.generate

The mixed path works, but it requires hand-rolled orchestration in application code. The repository's examples/full-pipeline/run.py follows the same manual pattern.

Rough edge

Customers with RAG pipelines likely expect a reusable full-pipeline runner that:

• calls retrieval
• passes retrieved context to generation
• records retrieval and generation per-fixture details together
• optionally runs reviewer predicates or judges
• aggregates both retrieval and generation metrics into one EvalResult/report card shape

Today they can build this manually, but the first integration has more glue than retrieval-only or generation-only paths.

Recommendation

Add either:

• a run_full_pipeline_category(...) helper, or
• a docs section explicitly naming the manual orchestration pattern as the supported v0.1 approach.