test: [YW-266] stabilize native-DB-concurrency test flakes (WsManager + mapping-store)

Two native-DB (PGlite/WASM) test races that ate CI re-runs across unrelated
PRs this cycle. Fixed at the mechanism level — no test.retry(), no bare
timeout bumps to mask the races.

server-core (mapping-store flake):
  vitest 4.x runs all five server-core test files in parallel forks. Every
  file spins up PGlite (WASM Postgres), so they cold-start the WASM module
  simultaneously and contend for file descriptors and memory on a loaded CI
  runner. The first test in each file — "mapping-store > get returns the
  record written by set" — was the most frequent victim because it runs
  before any warm-up query. Serialize the suite (maxWorkers/minWorkers 1) so
  the files cold-start one at a time. The tests are I/O-bound, so the
  wall-clock cost is small.

client / WsManager (the original report) is fixed in the wystack submodule
  bump: ws.test.ts leaked ~12 PGlite WASM instances (sync afterEach never
  closed them), squeezing the 2500ms-sleep no-reconnect tests against the 5s
  bun timeout. The submodule commit adds a tracked-handle registry drained in
  an async afterEach plus timeout headroom. See wystack PR for detail.

De-flake verified locally, each file run 20x with 0 failures:
  - server-core mapping-store.test.ts: 20/20 pass
  - client ws.test.ts (submodule): 20/20 pass

NOTE: submodule points at the wystack YW-266 branch commit; re-point to the
merged wystack main commit before merging this PR.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: youhaowei

libs/wystack

@@ -12,5 +12,15 @@ export default defineConfig({
     // tolerates a slow runner without masking a genuine hang.
     testTimeout: 60_000,
     include: ["**/*.test.{ts,tsx}"],
+    // All five test files in this package use PGlite (WASM Postgres). Running
+    // them in parallel forks causes simultaneous WASM cold-starts that contend
+    // for file descriptors and memory on a loaded CI runner — the first test in
+    // each file (mapping-store: "get returns the record written by set") was the
+    // most common victim because it runs before any warm-up query. Serialising
+    // to a single fork eliminates the cold-start race at the cost of slightly
+    // longer wall-clock time (the tests are I/O-bound, not CPU-bound, so the
+    // loss is small).
+    maxWorkers: 1,
+    minWorkers: 1,
   },
 });