feat: chunk-aware hybrid search in all text adapters (#854)

[pyramation]

Summary

Implements chunk-aware text search across all 4 search adapters (tsvector, BM25, trgm, pgvector) via the @hasChunks smart tag (#854). This enables hybrid search — applications can query both vector and text indexes on chunks tables simultaneously, then combine results with RRF or other fusion strategies.

New shared module — adapters/chunks.ts:

• ChunksInfo interface with all chunk metadata fields: chunksTable, parentFk, parentPk, embeddingField, contentField, searchField, searchIndexes
• getChunksInfo(codec) — extracts and validates chunk metadata from @hasChunks smart tag
• Handles both object and JSON-string tag formats

Adapter changes:

Adapter	Chunk query pattern	Score combinator	includeChunks option
tsvector	(SELECT MAX(ts_rank(search, tsquery)) FROM chunks WHERE fk = parent.id)	GREATEST(parent, chunk) — higher is better	✅ default true
BM25	(SELECT MIN(score <@> bm25query) FROM chunks WHERE fk = parent.id)	LEAST(parent, chunk) — lower is better	✅ default true
trgm	(SELECT MAX(similarity(content, value)) FROM chunks WHERE fk = parent.id)	GREATEST(parent, chunk) — higher is better	✅ default true
pgvector	Already had chunk support — refactored to use shared ChunksInfo	LEAST(parent, chunk) — lower is better	✅ (existing)

Each adapter checks searchIndexes from @hasChunks to verify the chunks table has the relevant index type before enabling chunk-aware querying.

Integration tests:

• finds parent via chunk tsvector match (term only in chunks) — verifies parent is found when search term only exists in chunk content
• finds parent via chunk trgm similarity (term only in chunks) — verifies fuzzy matching through chunks
• Setup SQL updated with content, search (tsvector), BM25, and trgm indexes on posts_chunks table

Also includes the prerequisite #856 work (search_indexes parameter on ProcessChunks/ProcessFileEmbedding).

Review & Testing Checklist for Human

Medium-high risk — core search infrastructure change affecting all 4 adapters.

• Verify the lateral subquery pattern generates correct SQL for each adapter by inspecting the generated queries (enable PostGraphile query logging)
• Test with a real provisioned database: create a table with ProcessChunks + search_indexes, insert parent + chunk rows, and query via each adapter
• Verify includeChunks: false correctly disables chunk querying for all text adapters (tsvector, BM25, trgm)
• Check performance: lateral subqueries on large chunks tables should use the appropriate indexes (GIN for tsvector/trgm, BM25 for pg_textsearch, HNSW for pgvector)
• Verify the BM25 chunks index name convention ({chunks_table}_{content_field}_bm25_idx) matches what the DB generator creates

Notes

• This PR includes the (WIP) Feat/v5 relations cont #856 commit (search_indexes parameter) since that work hasn't been merged to main yet
• The companion constructive-db PR (#1165) expands the @hasChunks smart tag with the search metadata these adapters consume
• The actual search orchestration (dual queries, RRF fusion) lives in application code — this PR enables that capability by making both text and vector search paths available on chunks

Link to Devin session: https://app.devin.ai/sessions/2b5a29d83d3f478e8d3d972653b4879c
Requested by: @pyramation

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring