fix(gateway): retry lifecycle notifications

[aoeman84]

Summary

• Preserve update completion marker files when final delivery fails so the next gateway startup can retry
• Preserve restart completion markers when adapters are not ready or sends fail
• Keep marker cleanup for successful delivery and intentional suppression cases

Tests

• /home/aoeman/.hermes/hermes-agent/venv/bin/python -m pytest -q tests/gateway/test_update_command.py tests/gateway/test_update_streaming.py tests/gateway/test_restart_notification.py

[teknium1]

Thanks for addressing a real delivery-loss path. The exception and restart-marker preservation idea still applies on current main, but the patch needs a targeted port.

Problems

• The update paths still do not inspect adapter.send() results. Adapters can return SendResult(success=False) rather than raise; current restart handling explicitly documents and checks that contract at gateway/run.py:14875-14886. The PR would still remove update markers after that kind of failed delivery.
• Current main now uses _thread_metadata_for_target and _non_conversational_metadata for these sends (gateway/run.py:14785-14810, 14862-14874). Preserve those current routing semantics when salvaging the retry behavior.

Suggested changes

• Port marker retention for exception and unsuccessful SendResult paths into current _send_update_notification and _watch_update_progress, while retaining the current metadata construction.
• Add SendResult(success=False) regression coverage for both update delivery paths; the existing restart test documents this adapter behavior at tests/gateway/test_restart_notification.py:490-524.

Automated hermes-sweeper review.