perf(postgres): optimize PgMemoryStore for throughput and query efficiency

[ZhangHuiGui]

Summary

Six targeted performance improvements to the PostgreSQL memory store (src/core/store/postgres.ts), reducing latency and improving throughput for high-frequency operations.

Changes

1. Batch Transaction Writes (upsertL1Batch / upsertL0Batch)

• Wrap multi-record inserts in a single BEGIN/COMMIT transaction
• Eliminates per-row connection acquisition and network round-trips
• Falls back to single-record path when batch size is 1

2. Hybrid Search SQL Rewrite (searchL1HybridNative)

• Carry all columns through CTEs, eliminating the final JOIN back to the main table
• Add WHERE content <@> to_bm25query(...) > 0 to exclude zero-relevance BM25 results from RRF fusion
• Use named prepared statement for server-side plan caching

3. Named Prepared Statements for High-Frequency Searches

• search_l1_vector_v1, search_l1_fts_v1
• search_l0_vector_v1, search_l0_fts_v1
• l1_hybrid_native_v2

Avoids repeated parse+plan overhead on every call.

4. Concurrent Reindexing (reindexAll)

• Replace serial embed+update loop with paginated reads
• Bounded-parallelism embedding (CONCURRENCY=4, Promise.allSettled)
• Batched UPDATE via unnest() arrays instead of per-row UPDATE

5. VACUUM ANALYZE After Bulk Deletes

• Fire-and-forget VACUUM ANALYZE when >50 rows are deleted
• Prevents DiskANN/HNSW index bloat from dead tuples

6. BM25 Query Normalization Improvement

• Add Chinese stopword filtering (60+ high-frequency function words)
• Filter single-character CJK tokens (too generic for BM25 precision)
• Cap token count at MAX_BM25_TOKENS=8 to maintain query precision

Impact

Area	Before	After
Batch write (10 records)	10 round-trips	1 transaction
Hybrid search	2 table scans (CTE + JOIN)	1 scan (CTE only)
Reindex 1000 records	~1000 sequential API calls	4-way parallel + batched UPDATE
Search plan caching	None (re-parse every call)	Server-side prepared statements

Testing

• Zero TypeScript lint errors after all changes
• Backward compatible: single-record paths unchanged
• Graceful degradation: all try/catch patterns preserved