hermes - 💡(How to fix) Fix bug(signal_rate_limit): TOCTOU race in get_scheduler() — concurrent calls can create duplicate SignalAttachmentScheduler instances

Summary

gateway/platforms/signal_rate_limit.py:get_scheduler() (line 351) uses a bare if _scheduler is None: guard with no lock. Gateway platforms dispatch concurrent coroutines, so two concurrent calls can both see _scheduler is None, both construct SignalAttachmentScheduler(), and the loser's instance is orphaned with its internal token-bucket state.

_scheduler: Optional[SignalAttachmentScheduler] = None  # no lock

def get_scheduler() -> SignalAttachmentScheduler:
    global _scheduler
    if _scheduler is None:              # ← race
        _scheduler = SignalAttachmentScheduler()  # ← duplicate instances

Impact

• Two scheduler instances means rate-limit tokens are not shared — each instance has its own bucket, effectively doubling the allowed burst rate.
• _reset_scheduler() (test helper) already shows the singleton is expected to be unique.
• No threading import in this file — needs to be added.

Fix

Add import threading and _scheduler_lock = threading.Lock(). Apply double-checked locking to get_scheduler(). Same pattern as #24731.