test: close active dev-up stub connections on signal

[Chris0Jeky]

Summary

• Stop accepting new connections before invoking server.close() in the synthetic Node frontend/API servers.
• Force-close active HTTP connections so a client that leaves a socket open cannot hold TERM teardown indefinitely.
• Add an exact ordering assertion on every platform and a live open-socket TERM regression on Linux/macOS, where the JavaScript signal listener is exercised.

Refs #2135 #2111 #2112

Verification

• node --check scripts/ci/dev-up.test.mjs
• node --test --test-concurrency=1 --test-name-pattern="launchers encode" scripts/ci/dev-up.test.mjs — 1 passed
• git diff --check

The full local launcher suite reached 75 passed before the existing Windows/Git-Bash instability forced the suite watchdog: 3 unrelated tests failed and 4 were cancelled on Bash timeout-shaped failures (the same #2378 cohort); all exercised PowerShell cases passed. The new POSIX-only live-socket test is not runnable on this Windows host and is intended for the Linux/macOS lanes.

NOT verified

• Hosted Linux/macOS execution of the new live-socket regression.
• The full hosted launcher matrix.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[Chris0Jeky]

Independent coordinator review at exact head 6c10850e688a3c51d67107e2bdfdad1ea2934cd2: no CRITICAL/HIGH correctness, security, or data-loss findings.

The changed helper ordering is safe for the Node 24.13.1 engine contract: it stops new accepts with server.close() and then clears active HTTP connections with closeAllConnections(). The cross-platform source assertion passes, and the POSIX-only live test is appropriately isolated from Windows native-signal behavior. The focused test is green. No fix requested.

The full local launcher suite's Bash failures are the pre-existing Windows/Git-Bash timeout cohort tracked under #2378/#2159/#2161, not failures in this diff; the corresponding PowerShell cases passed.

[Chris0Jeky]

Review (agent half of the ADR-0066 gate; Codex credits exhausted, SC-9): one fresh-context reviewer confirmed the stub teardown cannot hang (server.close then closeAllConnections; Node 24 pinned) or double-resolve (single process.exit path), the Linux/macOS regression is non-vacuous (an accepted-but-silent socket would keep close pending under the old helper) and cleans up in finally, and the launcher scripts are untouched. Verdict SHIP. MEDIUM worth a one-line change or follow-up: the readiness poll at ~1063 uses canBind, which LISTENS on the port every 50 ms while the child is racing to bind it — in success mode the helper rethrows EADDRINUSE, so a required-gate flake vector with no diagnostics (stdio ignore); poll with net.createConnection retrying on ECONNREFUSED instead. LOWs: the source-string ordering assertion is self-referential; the message overstates an ordering invariant; 10 s budget is the tightest in the file; stderr discarded. Overlap with parked #2522: helperSource stop path (~142-159) and the insertion point before the platform loop (~1033-1092). CI-control region: parked for the maintainer's review per ADR-0066 / SC-10.