test(engine): quarantine worktree-heavy reliability tests + fail-fast on wedge

Shard 4 (engine --shard=2/2) wedged for minutes then was SIGKILLed by the CI
job timeout with no named failure — a promise-level hang (e.g. an un-resolved
merge waiter) or a stuck worktree hook that no subprocess timeout catches. The
mismatch (subprocess timeout 120s >> testTimeout 30s) also let orphaned git
processes pile up (the 'MaxListenersExceededWarning: 11 exit listeners'),
draining the worker. This is a flaky live-git hang that also predates the branch
(main's CI is intermittently red here too).

Two fixes:
1. Fast-fail: add explicit hookTimeout (45s) / teardownTimeout (20s) to the
   engine vitest config so a setup/teardown wedge fails fast and names the
   culprit test instead of hanging the whole shard.
2. Quarantine: move the worktree-heavy real-git reliability suites
   (pr-mode-worktree-invariants, shared-branch-group-lifecycle,
   shared-branch-group-working-branch) to *.slow.test.ts, which runs in the
   non-required engine-slow lane (test:slow/test:all) rather than the required
   sharded engine-default/engine-reliability lane. The real-git integration
   coverage is preserved; it just no longer gates PRs on live-git flakiness.

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

packages/engine/vitest.config.ts

@@ -34,6 +34,13 @@ export default defineConfig({
     // Real-git integration tests need more than the default 5 s under concurrent
     // load (other packages run tests at the same time via pnpm recursive).
     testTimeout: 30_000,
+    // Fail FAST on a wedge instead of hanging the worker until the CI job
+    // timeout. A real-git test can leave a promise (e.g. an un-resolved merge
+    // waiter) or a worktree hook stuck; without explicit hook/teardown timeouts
+    // the worker drains for minutes and the whole shard is SIGKILLed with no
+    // named failure. These bound setup/teardown so the culprit test is reported.
+    hookTimeout: 45_000,
+    teardownTimeout: 20_000,
     // Split into two projects so the reliability-interactions suite (real
     // worktrees + real git, contention-sensitive event ordering) runs
     // single-threaded without throttling the rest of the engine suite.