fix(gateway): serialize startup auto-resume before inbound

[teknium1]

Summary

Gateway startup now queues real inbound messages until restart-interrupted auto-resume turns have been scheduled and completed, preventing duplicate agents for the same session after a restart.

Changes

• gateway/run.py: adds a startup restore gate, tracks startup auto-resume tasks, waits for their adapter-level processing tasks, then drains queued inbound messages in order.
• tests/gateway/test_restart_resume_pending.py: covers real inbound queuing during restore and verifies queued messages replay only after the synthetic resume turn finishes.

Validation

Check	Result
py_compile gateway/run.py tests/gateway/test_restart_resume_pending.py	passed
scripts/run_tests.sh tests/gateway/test_restart_resume_pending.py	70 passed
scripts/run_tests.sh tests/gateway/test_platform_reconnect.py tests/gateway/test_restart_resume_pending.py	100 passed

Fixes #45682.

Infographic

[liuhao1024]

Verification: LGTM — clean startup-restore serialization

Reviewed the full diff. The race condition is clear: platform adapters start receiving inbound messages as soon as they connect, but restart-interrupted sessions aren't auto-resumed until _schedule_resume_pending_sessions() runs. A user message arriving during the gap competes with the synthetic resume turn for the same session.

The fix introduces a startup-restore gate that:

1. Sets _startup_restore_in_progress = True before adapter connection
2. Queues non-internal inbound messages in _handle_message while the gate is active
3. Waits for all resume tasks to finish via _finish_startup_restore()
4. Drains the queue before releasing the gate

The asyncio.shield(task) in _run_startup_resume_event correctly waits for the agent turn to complete (not just handle_message returning after spawning the background task). The _hermes_startup_restore_replay marker prevents drain-replayed events from being re-queued.

Test coverage is solid — two async tests covering the gate-queue and wait-then-drain scenarios. The early-exit path (startup failure → _startup_restore_in_progress = False) is also covered in the start() method's error branches.

Also bundled: Telegram rich messages default-on (removes _rich_messages_enabled opt-in gate) + updated platform hint. This is a behavioral change but well-scoped — the rich path already had the latching mechanism (_rich_send_disabled) to handle unsupported bot versions.

[github-actions]

🔎 Lint report: hermes/hermes-b4d8e026 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10894 on HEAD, 10893 on base (🆕 +1)

🆕 New issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/gateway/test_restart_resume_pending.py:1245: [invalid-assignment] invalid-assignment: Object of type `def fake_handle_message(event: MessageEvent) -> CoroutineType[Any, Any, None]` is not assignable to attribute `handle_message` of type `def handle_message(self, event: MessageEvent) -> CoroutineType[Any, Any, None]`

✅ Fixed issues: none

Unchanged: 5725 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.