feat(retrieval): add RAG eval improvement loop

[drewstone]

What changed

• Added runRetrievalImprovementLoop() for one-call RAG/retrieval config optimization on top of agent-eval.
• Added retrieval eval scenarios, deterministic retrieval metrics, cost reporting, parameter search helpers, and root exports.
• Added SOTA-informed RAG eval completion roadmap covering retrieval, context quality, answer quality, diagnosis, and production gates.
• Updated README and architecture docs to reflect agent-knowledge as a knowledge-base construction and retrieval/RAG eval package.

Validation

• pnpm exec tsc --noEmit
• pnpm exec vitest run tests/retrieval-eval.test.ts
• pnpm exec tsup
• git merge-tree --write-tree origin/main HEAD

Notes

This PR was cut from a clean worktree based on origin/main because the original working branch had unrelated staged memory-holdout work and a pre-existing merge conflict against main.

[tangletools]

✅ Auto-approved drewstone PR — 3b3c4a9c

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-07T20:05:50Z}