test(daemon): repeated endpoint churn against real sockets

[BunsDev]

Third increment of cave-58eoq.4. Covers the bead's "repeated lifecycle stress detects orphaned processes, stale endpoint state, unsafe retries, and false healthy/fixed diagnostics" criterion.

Scoped narrow on purpose

Two neighbouring criteria are already met, and padding them would be noise rather than coverage:

• sleep/wake is pinned by daemon-connection-supervisor.test.ts — 501 lines, 14 tests, including hiding clears timers and aborts active work without new publication or backoff and foregrounding a started supervisor triggers exactly one immediate fresh request, plus the full backoff ladder and jitter bounds.
• single-transition endpoint faults are pinned by daemon-endpoint-faults.test.ts from increment 2 (test(daemon): endpoint fault coverage against real sockets #4500).

What was actually missing

The repeated case. A single bind/close cycle can pass while the tenth fails, and that is where the real failures live:

• an address that reports occupied after its owner left
• a listener that cannot rebind because something held the path
• a probe that leaks one connection per cycle and slowly starves the daemon of accept slots

Each of those reaches a user as "the daemon won't start" or as a duplicate daemon — and none reproduces on the first try, which is exactly why single-shot tests miss them.

The four tests

test	what it catches
12 bind/release cycles	a transition that drifts — stale occupied after release, or a cycle that stops reading free
rebind after churn	EADDRINUSE on the next launch because a released path was left behind
concurrent probes during churn	a launcher, status poll and recovery banner disagreeing — one stale occupied wrongly refuses a launch
per-cycle orphan check	a probe that holds its connection open, accumulating one per cycle

Addresses are platform-appropriate — a named pipe on Windows, a temp socket elsewhere — so the same cases run on both legs.

Verification

4 tests in 140ms; check:tests-wired (1607 files), pnpm typecheck, and pnpm lint all exit 0.

Remaining on the parent bead

Real-daemon (process-level) startup/shutdown stress, which needs spawning the actual binary rather than a socket stand-in. Per-PR macOS coverage is cave-5tpxy.

Built in the existing cave-58eoq.4.1 worktree rather than creating another — the checkout is at 31 against a 28 budget with writer-intent contention, and adding to that to avoid a git checkout -b would have been the wrong trade.

[Copilot]

Pull request overview

Adds a new daemon connectivity test suite that repeatedly churns real socket/pipe endpoints to catch lifecycle drift issues that single-transition tests can miss (stale “occupied” state, rebind failures, and probe-connection residue).

Changes:

• Introduces daemon-endpoint-churn.test.ts with 4 churn-focused tests exercising the real connect({ path }) probe across multiple bind/release cycles.
• Wires the new test file into the test:app/test:api/test:mobile runner via scripts/run-tests.mjs.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/lib/daemon-endpoint-churn.test.ts	Adds repeated endpoint churn tests over real sockets/named pipes to detect drift and connection residue across cycles.
scripts/run-tests.mjs	Registers the new churn test in SUITES so CI picks it up.

Suppressed comments (1)

src/lib/daemon-endpoint-churn.test.ts:155

• In the orphan/leak test, the per-cycle server is not added to cleanups. If the test fails mid-cycle (e.g. unexpected rejection from inspectDaemonAddress/unbind), the listener can remain open and leak resources. Add each created server to cleanups immediately after construction.

  for (let cycle = 1; cycle <= CYCLES; cycle += 1) {
    const server = createServer(() => {});
    server.on("connection", (socket) => {
      everOpened += 1;
      stillOpen += 1;

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[BunsDev]

Fixed in 01b883f461, and the review was right.

bind() created a real server without registering it in the file-level cleanups list. These tests assert mid-cycle, between bind() and unbind(), so an assertion that throws skips its unbind and leaves a listener holding the address — one genuine failure would cascade into a hung suite. That is exactly the behaviour the fault harness in #4489 exists to prevent, and increment 2's listenOn() already registers correctly; this file did not.

One addition beyond the reported line: the orphan-check test creates its own server inline (it counts connections, so it cannot use the helper) and had the identical flaw. Both are fixed.