Fix: add lock-protected reads on _sync_failures dict in scheduler

[neuromechanist]

Summary

src/api/scheduler.py uses a shared mutable dict _sync_failures across threads (background scheduler jobs run in separate threads via APScheduler). Reads of _sync_failures[key] at line 57 occur outside the lock, creating a race condition with concurrent writes.

Location

• src/api/scheduler.py:52-66: _track_failure() writes under lock but reads at line 57 (count = _sync_failures[key]) happen without holding _sync_failures_lock.
• Other read sites: _reset_failure() (line 80, pop), _cleanup_mirrors (lines 301-303), _check_community_budgets (lines 361-379).

Fix

Wrap all reads of _sync_failures under the same lock used for writes:

with _sync_failures_lock:
    count = _sync_failures.get(key, 0)

Acceptance Criteria

• All reads of _sync_failures are protected by _sync_failures_lock
• New unit test exercises concurrent read/write pattern
• No race conditions detected under thread sanitizer (if available)