Thanks for this well-structured PR, @runlvl — the problem you identified is real and the fixes are solid. However, all four concerns addressed here have since been independently resolved on main.

Automated hermes-sweeper review found the following implementations already merged:

• Session continuity across drain-timeout restarts — cb4addaca (PR spec: automatic session resume after gateway restart #11852, 2026-04-18) introduces SessionEntry.resume_pending and marks still-running sessions resume_pending before interrupting on drain timeout. suspend_recently_active() skips resume_pending entries so they auto-resume on next message. See gateway/session.py:461 and gateway/run.py:2531.

• Clean-shutdown marker / exclude planned-restart sessions from startup auto-suspend — b6b6b02f0 (PR fix: prevent unwanted session auto-reset after graceful gateway restarts #8299, 2026-04-12) writes .clean_shutdown on graceful stop and skips suspend_recently_active() on the next startup when the marker is present. tests/gateway/test_clean_shutdown_marker.py (226 lines) is already on main.

• Prefer persisted session origin for restart/shutdown notifications — 0613f10de (2026-04-20) makes _notify_active_sessions_of_shutdown() read session_store._entries[key].origin first, falling back to _parse_session_key() for legacy/test sessions. Fixes colon-ID misrouting (e.g. Matrix room IDs). See gateway/run.py:1571.

• Telegram BadRequest treated as permanent failure — 41d9d0807 (PR fix(telegram): fall back to no thread_id on 'Message thread not found' #3390, 2026-03-27) detects BadRequest inside the NetworkError handler: thread-not-found retries without thread_id; all other BadRequest errors (including Chat not found) raise immediately and are not retried. Covered by tests/gateway/test_telegram_thread_fallback.py. See gateway/platforms/telegram.py:1053.

• Stale restart notify targets dropped — _send_restart_notification() (gateway/run.py:8031) already returns early when the adapter is not connected and always unlinks the notify file in finally, so stale targets are dropped without retry on any send failure.

Closing as implemented_on_main. If the e8b21979 commit adds more granular explicit logging for Chat not found inside _send_restart_notification specifically, that's a narrower style improvement worth a fresh focused PR.

Sorry for that PR, my Hermes Agent instance opened it on his own. To avoid this in the future I gave him a global rule to always check if there are already covering PRs or even a fix before opening a new PR.
EDIT: I added an LLM labeling to this global rule too, so that it is transparent that an LLM opened the PR.