* fix(bin): work around libuv async assertion on Windows exit

trueconf-setup crashes at the end of a successful run on Windows + Node
24.x with `assertion failed !(handle->flags & UV_HANDLE_CLOSING)` in
src/win/async.c. Root cause is the open Node issue nodejs/node#56645 —
Node calls `uv_async_send()` on a closing handle when `process.exit()`
short-circuits libuv teardown right after `fetch()`. PR
nodejs/node#61999 has been open since Feb 2026 with no backport.

Two cooperating workarounds, both documented in the upstream thread:

- src/probe.mjs: validateOAuthCredentials now always owns its undici
  Agent and closes it in finally, so no keep-alive sockets dangle into
  loop teardown. Cleanup is best-effort — errors are swallowed so the
  OAuth result classification stays unaffected.

- bin/trueconf-setup.mjs: CLI entry sets `process.exitCode` instead of
  calling `process.exit()`, letting the loop drain naturally after the
  dispatcher and clack readline release their handles. Shell exit
  contract is unchanged.

Tests:
- probe.test.mjs: assert dispatcher.close() runs on the success path
  and on the fetch-throws path.
- bin-cli-subprocess-exits.test.ts: spawn the CLI as a real subprocess
  and assert it exits within an 8s budget on success (0) and OAuth
  failure (1) — catches loop-hang regressions on any platform.

970 tests | 2 skipped on macOS, tsc clean, oxlint clean. Windows UAT
pending.

* fix(bin): watchdog + tighter exit-budget test on review feedback

Two review findings on the prior libuv-assertion commit:

- bin/trueconf-setup.mjs: a future regression that leaks an event-loop
  handle (clack stdin in raw mode, sharp libvips worker, an unref-missed
  timer) would now silently freeze the wizard's terminal forever instead
  of crashing. Add a 10-second .unref()'d watchdog that prints a clear
  diagnostic to stderr and forces exit, so a leak is loud, not silent.
  The timer self-defeats on the happy path because .unref() lets the
  loop drain ahead of it.

- tests/integration/bin-cli-subprocess-exits.test.ts: the prior version
  had three weaknesses. (1) The CLI entry block does not parse
  --config from argv, so passing `--config /tmp/x.json` was silently
  ignored and the success test wrote a fake-server config into the
  developer's real ~/.openclaw/openclaw.json. Sandbox via `HOME` and
  `USERPROFILE` env vars instead — that's the only knob the wizard's
  default `homedir()` lookup reads. (2) The 8s budget was too lax: a
  partial leak (5s timer ref) would slip under the kill timer
  unnoticed. Assert elapsedMs < 3000 explicitly so partial regressions
  fail loud. (3) Hard-coded port `1` for the failure path can SYN-drop
  on CI firewalls and time out, masking a real loop-hang as a flake.
  Reserve a closed loopback port via listen(0) → close instead — that
  reliably ECONNREFUSEs.

970 tests | 2 skipped on macOS, tsc clean, oxlint clean. Windows UAT
still pending.