Summary

• track background message-processing tasks spawned by platform adapters
• interrupt running agents and cancel adapter background tasks during gateway shutdown before adapters disconnect
• clear shutdown-time pending session state and add regression coverage for restart/shutdown behavior

What this addresses

Issue #1414 reports that after stopping a busy gateway and restarting with hermes gateway run --replace, the old task can appear to keep going, task/progress labels can flicker, and the restarted gateway can fall into a bad state while the previous in-flight work is still unwinding.

I did not reproduce the exact OpenRouter 502 sequence deterministically, but I did isolate a concrete shutdown bug on current main:

• platform adapters spawn background message-processing tasks and do not track them
• GatewayRunner.stop() disconnects adapters but does not cancel those tasks
• GatewayRunner.stop() also does not interrupt agents already recorded in _running_agents

That means an old gateway instance can keep working on in-flight message tasks during shutdown/replacement instead of being cleanly quiesced first.

Test plan

• source .venv/bin/activate && python -m pytest tests/gateway/test_gateway_shutdown.py -n0 -q
• source .venv/bin/activate && python -m pytest tests/gateway/test_gateway_shutdown.py tests/gateway/test_interrupt_key_match.py tests/gateway/test_telegram_documents.py tests/gateway/test_telegram_photo_interrupts.py -n0 -q
• source .venv/bin/activate && python -m pytest tests/gateway/ tests/hermes_cli/test_gateway.py -n0 -q