Retire the evaluation spike after coverage parity

What to build

Retire the experimental rhdh-templates evaluation spike after proving that permanent evaluation sources cover its useful scenarios and that historical decisions remain discoverable.

Acceptance criteria

• A coverage comparison maps every retained spike scenario to a permanent behavior, routing, or uplift case and identifies anything intentionally excluded.
• Any AEH-generated case counted toward coverage parity was reviewed and promoted to a permanent case; unreviewed candidate output is excluded.
• Before spike retirement, at least one paired uplift pilot is rerun with MLflow enabled and its model/configuration metadata, per-case scores, artifacts, and traces are reviewed; MLflow's local database and artifacts remain uncommitted.
• Historical framework findings and artifact-policy rationale remain available in maintained documentation before spike files are removed.
• Spike removal is isolated from unrelated refactoring, and permanent local suites plus the repository test suite still pass afterward.
• Raw pilot artifacts, credentials, and local MLflow data are not added while retiring the spike.

Blocked by

• #15

Start by reviewing blocked issue #15 and the retained evaluation scenarios. Map each scenario to permanent coverage, review AEH-generated cases, and preserve the framework findings and artifact-policy rationale in maintained documentation. Before removal, rerun the paired uplift pilot with MLflow enabled, then run the permanent local suites and repository test suite; done means the spike is removed without committing artifacts or local MLflow data.