perf: add revision-pinned retrieval serving index

[suguanYang]

Summary

• Add revision-pinned retrieval serving manifests, statistics, and namespace generations.
• Use one coherent revision snapshot through classic/map-nav scoring, ranking, hydration, references, and connected assets.
• Add persisted map-unit indexes, trigram term filtering, concurrent classic channels, duplicate-safe RRF, and request/episode-local reuse.
• Guard publication, archive, namespace moves, and backfill races against stale serving data.
• Add contract coverage, ADRs, the retrieval-serving plan, and a dedicated DevOps operations runbook.

Validation

• make check (Ruff clean; Pyright 0 errors).
• Focused retrieval contracts pass.
• Local replay against a copied production-shaped corpus: classic request HTTP 200 in ~3.5s; default map-nav request HTTP 200 and completed with persisted serving indexes.
• No prompts were changed.

DevOps operations

docs/design/retrieval-serving-index-plan.md now has a dedicated runbook covering:

1. Read-only preflight inventory and rollback point.
2. Additive migration order.
3. Low-traffic trigram index rollout and lock monitoring.
4. Controlled, bounded, idempotent active-revision backfill.
5. Manifest/checksum/count/generation reconciliation.
6. Legacy-versus-serving differential acceptance checks.
7. Safe rollback and repair behavior.

Production migrations and backfill require explicit approval immediately before execution. Incomplete or invalid serving data falls back to the exact legacy reader.