[High] Memory layer N+1 elimination — recall UPDATE + vector hydrate

Severity: High · Domain: Performance
Status: Merges 1 earlier issue into a single tracking item.

Sub-findings rollup

Why merged

Both N+1s are in the same file (semantic.rs) on the same call chain — a single PR can eliminate both most economically.

Combined fix plan

1. Wrap recall UPDATEs in a transaction:

   let tx = conn.transaction()?;
   let mut stmt = tx.prepare("UPDATE fragments SET last_used=?1, hit_count=hit_count+1 WHERE id=?2")?;
   for f in &recalled { stmt.execute(params![now, f.id])?; }
   tx.commit()?;

   Amortizes WAL fsync over a single transaction.

2. Convert vector hydrate to a batched IN query:

   let placeholders = (0..ids.len()).map(|_| "?").collect::<Vec<_>>().join(",");
   let sql = format!("SELECT id, ... FROM fragments WHERE id IN ({placeholders})");

   Single SELECT; for K=50, drops 50 round-trips to 1.

Tests

• Unit test: recall of 100 fragments → write latency ≤ single WAL fsync × 2.
• Unit test: vector ANN with K=50 → SQL call count = 1 (no longer grows with K).
• Bench: end-to-end recall of 100 fragments improves ≥ 10× over baseline.