[datlechin]

Why

Release criterion 4 still needed the 1,000,000-row page-and-export test on SQLite. Its suites run in the unit suite under cargo test, where one binary's tests share a process, so a resident peak measured there would count theirs too.

Change

• crates/drivers/sqlite/tests/large_table.rs is a test target holding only this case, so its process is its own. The fast job's unit suite runs it.
• large_table_conformance_suite! takes docker: true | false. The five server drivers pass true.
• SQLite's adapter moves to tests/integration/sqlite_engine.rs, which both targets include. It fills the table with a recursive CTE.
• The roadmap row and testing docs say SQLite runs it.

Verified

• Docker, debug: walk 3.6 to 12 s, export 3.5 to 9.2 s, run 9 to 22 s. Release: about 5 s. Peak growth 2 MiB or less. The shared bounds hold.
• With stream_read buffering every row, the case fails at +313.7 MiB.
• Adds about 10 to 25 s of run and one test binary to build per unit-suite job.