fix(recall): exclude self-match and make ef_search pooler-safe

[geekyfox90]

What & why

Two correctness fixes in the recall path, found during a review of the benchmark methodology.

1. Self-match inflates recall@k

sampleQueryVectors draws query vectors straight from the indexed table, so each query's nearest neighbour is the row itself (distance 0). That self-match is a guaranteed hit in both the index result and the exact ground truth, so recall@k always includes one "free" hit (a useless index still scores ≥1/k; a real 7-of-9 shows as 8/10).

Fix: sampleQueryVectors now also returns each vector's ctid. Recall fetches k+1 and drops the self ctid from both the index result and the ground truth, so we measure the k real neighbours.

2. ef_search silently ignored behind a transaction pooler

The ef_search sweep applied a session-level SET hnsw.ef_search on otherwise-autocommit queries. Through a transaction pooler (PgBouncer / the Supabase pooler), consecutive statements can land on different backends, so the SET doesn't persist — the whole --ef-search sweep silently runs at the server default (identical rows).

Fix: each ef level now runs inside one transaction using SET LOCAL hnsw.ef_search (one BEGIN/COMMIT per level — no per-query overhead). This is pinned and honoured on both direct and pooled connections. (Matches the pattern the ground-truth seq-scan path already uses.)

Notes

• Latency/throughput are unchanged — they run at the default ef, so the hot timing paths are untouched.
• Builds clean, go vet clean. Removed a now-dead scanIDsConn helper.

🤖 Generated with Claude Code