fix(net): schedule close events independently of user timers

[steipete]

Socket close notifications now use Bun's native immediate scheduler even when user code or a fake clock replaces global setImmediate or node:timers.setImmediate.

closeSocketHandle previously closed the native handle and then looked up ambient setImmediate to emit close. A replacement that never runs its callback strands the event after the peer has already sent FIN. Bind node:net's existing immediate continuations directly to the native timer implementation. Keeping the check-phase scheduler, rather than moving close into nextTick, preserves the ordering that Node exposes relative to end, finish, nextTick and real immediates. The same internal binding covers reset and adopted-TLS close continuations.

Validation: four subprocess variants poison timers before/after importing net and cover peer/local initiation. All pass patched (317 ms including the test runner); both test groups fail at the explicit deadline on baseline e8105cb. Identical standalone probes pass on Node 24.21.0 and miss close in all four baseline cases. Net: 268 passed / 37 skipped. TLS: 457 passed / 32 skipped / 6 TODO. HTTP: 1,662 passed / 29 skipped / 3 TODO / 1 pre-existing Node-reference failure documented in #38. Source TypeScript, lint, formatting and independent P2 autoreview pass. No Rust changes. Rebased onto the separately merged TLS fix at a0ef910: only additive changelog entries conflicted; runtime/test/docs patch-id matches the tested original. Rebase review is clean; GitHub Lint and Format pass on d2d3657.

W15 measured 119.176 excess test-seconds in the old admin handler file and modeled 20–44 seconds of whole-config invocation opportunity. Those scheduling estimates are not measured patched speedups. Current OpenClaw main 95ed4ee478cd8, full admin handler file: baseline 12 passed / 1 timeout in 72.403 s; patched 13 passed in 46.040 s. Runs use one worker, private 0700 TMPDIR, the selected Bun first on PATH, OPENCLAW_VITEST_RUNTIME=bun, disabled disposable compile caching, 15 s test/hook deadlines and a 300 s outer guard. The 408 response and real client close both complete without advancing fake time further. Shared-host measurements are not isolated throughput benchmarks.

Upstream status: searches of oven-sh/bun issues and PRs for socket close/setImmediate and fake timers found no matching fix to port.