fix: preserve gateway update completion notifications

[Yellowfish23]

Summary

• prevent duplicate gateway /update runs from overwriting lifecycle markers
• preserve completion markers when final notification delivery transiently fails
• retry final update notifications and clean stale/corrupt marker files
• keep safe-update transcript output while preserving the update process exit code

Test Plan

• .venv/Scripts/python.exe -m pytest tests/gateway/test_update_command.py tests/gateway/test_update_streaming.py -q --timeout-method=thread

[alt-glitch]

Competes with open #30123 (retry lifecycle notifications) and #15539 (prevent concurrent gateway update clobbering).

[teknium1]

Thanks for addressing a real update-notification delivery-loss path. Current main still deletes update markers after a final send exception (gateway/run.py:14614-14623) and ignores unsuccessful SendResult values in the legacy completion path (gateway/run.py:14806-14825).

Problems

• The PR's streaming final-send hunk retries raised exceptions, but it does not inspect adapter.send() for SendResult(success=False). Current lifecycle code documents that adapters can return this instead of raising at gateway/run.py:14875-14886; the streaming path would still clean up undelivered completion markers.
• The PR changes _handle_update_command in gateway/run.py, but current main moved that handler to gateway/slash_commands.py:4504 in 619bd7827. The handler-side sentinel/transcript/UTF-8 changes need a targeted port that preserves current routing metadata.

Suggested changes

• Apply the same unsuccessful-SendResult retention logic in both _watch_update_progress and _send_update_notification, with regression tests for each path.
• Port the command-start changes into gateway/slash_commands.py and preserve current _thread_metadata_for_target / _non_conversational_metadata delivery behavior.

Automated hermes-sweeper review.