fix: await worker-thread termination on stop/close (fixes in-process WASM teardown crash)

[wmadden-electric]

Problem

Both worker pools tear down like this:

for (const w of this.workers) {
  w.worker.postMessage({ type: "stop" });
  w.worker.terminate();   // promise never awaited
}

worker.terminate() returns a promise that isn't awaited, so stop() — and the local-server close() path that calls it — resolves while the worker threads are still tearing down.

That's invisible for the standalone server (its own process), but @prisma/streams-local also runs in-process inside @prisma/dev. There, await localServer.close() returns while the touch/segmenter worker threads are still alive, and @prisma/dev immediately tears down PGlite (Postgres compiled to WebAssembly). A worker thread terminating at the same moment another thread frees PGlite's WASM JIT pages races V8's process-global JIT bookkeeping (PKU / ThreadIsolation) and aborts the whole process on Linux:

# Check failed: jit_page_->allocations_.erase(addr) == 1

Every operation succeeds; the process just dies during teardown. Downstream (prisma-next) it surfaced as a flaky vitest "Worker exited unexpectedly" with all tests passing — frequent in high-churn suites that create and close many dev servers in one process, rare in low-churn ones.

Fix

Await termination, and thread that await all the way up the close path so a resolved close() guarantees the worker threads are gone:

• TouchProcessorWorkerPool.stop() / restart() → async, await Promise.all(workers.map(w => w.terminate()))
• SegmenterWorkerPool.stop() → same
• TouchProcessorManager.stop() → async, awaits pool.stop() (and the in-loop pool.restart())
• app_core close() → async, awaits the worker-pool stops before the rest
• local/server.ts and server.ts → await app.close()

Notes

• DurableStreamsLocalServer.close() was already Promise<void>, so this is backward-compatible for callers that await it (@prisma/dev does).
• Verified: tsc --noEmit clean; local_server, segmenter_behavior, and touch_memory_journal suites pass. The crash itself only reproduces on Linux with an in-process WASM host (PGlite), which this repo doesn't have, so there's no repro test here — the prisma-next e2e suite is the end-to-end guard once a new @prisma/streams-local is published.

🤖 Generated with Claude Code