fix: bound the PGlite open so a stalled init can't wedge a turn

[davidj4tech]

Follow-up to #5, at the opposite end of the same lifecycle. #5 guards the close (headless exit hang). This guards the open.

The failure

Both index modules cache their in-flight open in a module-level initPromise. That cache turns one stalled open into a permanent hang: PGlite is a single-writer WASM Postgres over a shared dataDir (~/.pi/mega-compact-vector), so a second pi process on the same dir can block indefinitely. The await never settles, the dead promise is cached, and every later caller awaits it. With no timers and no sockets left, node reports "Promise resolution is still pending but the event loop has already resolved" — and the pi turn awaiting it never ends.

Observed in the wild: a session sat on one turn for 75+ minutes, all threads parked in epoll_wait/futex_wait, zero open sockets, PGlite files held open. Not slow — deadlocked. #5's fix cannot help here: the session never reaches session_shutdown.

The change

withOpenTimeout() puts a ceiling on the open — 30s default, configurable via MEGACOMPACT_PGLITE_OPEN_TIMEOUT_MS, 0 restores the previous unbounded behavior. On timeout the caller gets undefined and both modules degrade to the synchronous scan they already fall back to. The cached initPromise is cleared rather than left dead, and the abandoned open is disowned — if it settles later the instance is closed, so an orphan can't hold the dataDir lock.

Rejections deliberately propagate unchanged, so the existing corrupt-dir detection (Aborted/RuntimeError → wipe + one retry) still runs. Only the timeout resolves to undefined.

Tests

src/store/pgOpenGuard.test.ts — 6 tests: never-settling open returns promptly, success passes through, rejection still reaches the corrupt-retry path, late-settling open is closed, 0 disables the guard, env parsing rejects junk.

Full sweep: 70 of 76 test files pass. The 6 that don't are pre-existing and unrelated — 5 hang on a harness() teardown leak (a zero-test run exits in 148ms), 1 fails because the @mongodb-js/zstd native addon isn't built. I confirmed src/memoryOps.test.js hangs identically with these changes stashed.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.