[sfloess]

Summary

This PR advances the existing REST-first implementation branch and removes direct in-process capability wiring from ingestion/search orchestration.

• Adds independently addressable REST contracts for chunking, embeddings, and indexing.
• Routes document ingestion through those contracts.
• Routes search candidate enumeration and query embedding through the index and embedding contracts.
• Uses ASGI-backed HTTP transport while capabilities are co-located, preserving HTTP routing, request validation, and JSON serialization.
• Adds explicit API contract coverage for the capability endpoints.
• Keeps the existing durable store, migration behavior, revision semantics, and public document/search contracts.
• Corrects architecture, REST contract, implementation-status, and capability migration documentation.

Boundaries and known limitations

No POC services or POC data are accessed or modified. Authentication, partition enforcement, a dedicated vector index, bounded retrieval, reranking, graph, memory-specific endpoints, and production deployment hardening remain future work. The current full chunk enumeration and JSON vector scan are not production-scale retrieval claims.

The legacy SQLite Index compatibility issue was resolved with an explicit boundary decision: Ensemble does not migrate, rebuild, or import historical CE/POC Index databases. Those remain untouched under their existing owner. Any future import must use an explicit API or export/import process.

Verification

• GitHub Actions run 41362550739: passed, including install and test steps.
• Earlier implementation commit run 41362407372: passed.

Do not merge until the diff has been reviewed.