[kyscott18]

Second of four stacked PRs, on #112: db transport → PGlite transport → sim DB as a transport with injected errors → PGlite for the sim.

• DatabaseOptions takes { kind: "pglite" }, an in-memory PGlite that lasts as long as the process. { url, ... } is unchanged (kind is optional for Postgres), so existing configs keep working.
• makePgliteTransport (@electric-sql/pglite 0.5.8, the version abipg already pins): a real PGlite transaction, no retries (notApplied is never set: there is no connection to lose), and a no-op advisory lock (only this process can reach an in-memory database).
• runTransaction is the Effect-to-promise bridge, taken out of the Postgres transaction so both transports share it; the commit-status check stays a Postgres-only hook. A failed statement still fails the transaction even if f recovers. On PGlite, a COMMIT after a failed statement silently rolls back and resolves.
• Drizzle is now the driver-neutral PgDatabase<PgQueryResultHKT>. selectMaxExecutionIndex selects from the subquery (max(max)::text) instead of reading db.execute, which returns { rows } on PGlite but an array on Bun SQL.

Transaction guard. PGlite has one session. Its query/transaction already queue behind an open transaction, so transactions from separate fibers can't overlap. The real hazard is a statement on db instead of tx inside f. On PGlite that waits on its own transaction forever; on Postgres it silently runs outside the transaction. makeDbClient runs f with an InTransaction reference, and query/transaction die when it is set, including in fibers that f forks.

Known issue, no workaround yet. PGlite reads a text value that starts with U+FEFF (a BOM) without it. Its protocol reader decodes each field with new TextDecoder("utf-8"), which strips a leading BOM by default; the stored bytes are correct. Still the case in 0.5.14. Documented on makePgliteTransport. The runtime never reads text columns back, and the sim stores no strings since #111, so nothing here is affected. db-query.test.ts stays on Postgres: pointed at PGlite, its BOM case fails and every other string passes.

Checked:

• typecheck, lint, and bun test with DATABASE_URL (84 pass);
• a throwaway guard check on both backends: db.query in f and in a forked fiber die, while tx.query and two fibers' concurrent transactions pass;
• sim seeds 0xe, 0x11, 0x41 pass on Postgres, the same as refactor(typewriter): split the database into a transport and a client #112.

🤖 Generated with Claude Code