Epic: Hybrid retrieval implementation over reconstructable indexes

[voidrot]

Summary

Implement hybrid retrieval as reconstructable secondary indexes layered on top of the markdown source-of-truth corpus.

Why this work is needed

The repository already has a planning anchor for hybrid retrieval, but there is no implementation backlog yet. This epic turns that design direction into executable issues.

Scope

• Implementation-oriented backend decision for lexical and vector retrieval.
• Embedding and index build pipeline.
• Integration into the retrieval SDK and MCP service.

Non-negotiable constraints

• Live markdown and frontmatter remain the source of truth.
• Search indexes are reconstructable and disposable.
• Visibility filtering remains enforced before results leave the retrieval boundary.

Out of scope

• Moving documents into Postgres.
• Making ChromaDB mandatory before evaluating a Postgres-first path.
• Replacing the existing deterministic retrieval path without fallback.

Exit criteria

• The backend choice and implementation plan are explicit.
• Index build and retrieval integration work are tracked under this epic.
• The epic is clearly connected to the earlier planning work.

Related

• Follow-up to Plan hybrid retrieval over markdown with BM25 and vector search #44.