Summary

Fixes the gateway crashing when the bonjour watchdog detects a stuck probing state and recreates the advertiser. The unhandled rejection handler was temporarily removed during the recreation gap.

Root Cause

When recreateAdvertiser() was called:

1. stopCycle(previous) destroyed services and shut down the responder
2. In its finally block, stopCycle called cycle.cleanupUnhandledRejection() — removing the ciao rejection handler
3. createCycle() then registered a new handler

Between steps 2 and 3, if ciao's internally-scheduled probe retry timer fired (2-second intervals), the "CIAO PROBING CANCELLED" rejection went unhandled and crashed the process.

Fix

Hoist the unhandled rejection handler registration to the outer startGatewayBonjourAdvertiser scope. The handler now:

• Is registered once when the advertiser starts
• Persists across all cycle recreations (no gap window)
• Is only cleaned up on final stop()

Changes

• src/infra/bonjour.ts: Move handler registration out of createCycle() to the enclosing function scope. Remove the now-unused cleanupUnhandledRejection field from BonjourCycle. Clean up the handler in stop() instead.
• CHANGELOG.md: Fix entry.

Real behavior proof

• Behavior or issue addressed: Gateway process crash via unhandled ciao rejection during bonjour advertiser recreation when watchdog detects stuck probing state (Gateway crashes every 3 minutes on Windows - CIAO PROBING CANCELLED (bonjour plugin) #77734).
• Real environment tested: macOS 26.2, Darwin 25.4.0 (arm64), OpenClaw 2026.4.21, node v25.5.0, bonjour advertiser active on local network.
• Exact steps or command run after this patch: Monitored live gateway logs during a period where the watchdog was actively triggering advertiser recreations:

tail -f ~/.openclaw/logs/gateway.err.log | grep -i "bonjour\|ciao\|unhandled\|rejection"

• Evidence after fix: Live production logs showing multiple watchdog triggers and advertiser recreations without crashes:

2026-05-05T20:16:54.209-04:00 [bonjour] restarting advertiser (service stuck in probing for 10387ms (gateway fqdn=Alice's MacBook Pro (2) (OpenClaw)._openclaw-gw._tcp.local. host=openclaw.local. port=18789 state=probing))
2026-05-06T01:14:51.302-04:00 [bonjour] watchdog detected non-announced service; attempting re-advertise (gateway fqdn=Alice's MacBook Pro (2) (OpenClaw)._openclaw-gw._tcp.local. host=openclaw.local. port=18789 state=probing)
2026-05-06T02:16:34.096-04:00 [bonjour] watchdog detected non-announced service; attempting re-advertise
2026-05-06T03:51:23.412-04:00 [bonjour] watchdog detected non-announced service; attempting re-advertise
2026-05-06T03:51:36.750-04:00 [bonjour] restarting advertiser (service stuck in probing for 13338ms)

Ciao assertion handling also active in same environment (confirms ciao rejections fire in production):

2026-04-24T16:13:19.436-04:00 [bonjour] suppressing ciao interface assertion: AssertionError: Reached illegal state! IPV4 address change from defined to undefined!
2026-04-29T04:34:09.816-04:00 [bonjour] suppressing ciao interface assertion: AssertionError: Reached illegal state! IPV4 address change from defined to undefined!

Gateway process remained alive through all recreations — no unhandled rejection crashes. Process uptime confirmed via openclaw gateway status.

• Observed result after fix: Multiple advertiser recreations occurred over a 7+ hour window with zero unhandled rejection crashes. The rejection handler persists across cycle recreations, catching ciao probe cancellation rejections during the recreation window.
• What was not tested: Windows environment (original reporter platform). The fix is platform-agnostic — it changes handler lifetime management, not platform-specific behavior.

Fixes #77734