[jasperblues]

Closes #430. The reference fix is liberation-data/drivine4j#631; the TypeScript port is liberation-data/drivine#646.

Cause

#420 turned off enable_seqscan, enable_hashjoin and enable_mergejoin around a compiled write's per-row loop on Postgres. An autoanalyze that lands just after a DELETE leaves the node table's statistics at about one row, while the table still holds many pages. In that state, the loop's per-row join of a CALL's rows is planned like this:

Nested Loop Left Join  (rows=100)
  Join Filter: (n8.nid = ((r7.item ->> 'v1'::text))::bigint)
  Rows Removed by Join Filter: 17356
  ->  Function Scan on jsonb_array_elements r7  (rows=100)
  ->  Materialize  (actual rows=17357)
        ->  Index Scan using drivine_node_pkey on drivine_node n8  (rows=1) (actual rows=17357)

That is a full, unparameterized scan of the primary-key index under a Materialize, so every row reads the whole table. MATCH (n:CyLoop) CALL (n) { SET n.y = n.i RETURN n.i AS j } RETURN count(j) took between 72 s and 142 s over 20,000 rows. It failed in 3 of about 4 CI runs, sync and async alike. On a test machine it happened 3 times in 400 runs, with auto_explain capturing the plan above.

The fix

enable_material joins _JOIN_METHODS. Each wrapped loop now also runs with materialization off, so it reads the index by the row's key, and the setting is put back with the other three.

• The compiled SQL: only the 38 Postgres procedures A compiled MERGE of many relationships is quadratic on Postgres once the edge table is vacuumed empty #420 already wraps change, and only in their planner lines (enable_material and cy_planner[4]). The other 1,058 Postgres statements are byte-identical.
• Unchanged: no golden statement or Cypher spec case changes.

Tests

• test_compiled_write_loop_time.py, async and sync: on Postgres, the test runs ANALYZE drivine_node right after its DELETE and turns autovacuum off on that table while the writes are timed (reset in finally). Without the guard, an autoanalyze landing mid-run sometimes reset the statistics. With the bad state forced, the test fails every time without the fix (about 98 s over 20,000 rows) and passes with it (about 0.65 s).
• test_postgres_nested_loop_plan.py and test_postgres_nested_loop_settings.py: they cover the fourth setting, including its restore when the loop fails.

Docs

docs/04-statements.md names enable_material.

Gate

• ruff check, ruff format --check, pyright and unasync --check pass.
• The unit suite passes.
• The Postgres integration suites pass, together with the sync schema suites, through the test lock.
• Two review passes ran; the second was clean.