A race condition exists in the optional report buffer feature between retention cleanup and normal transcript writes. In `report buffer.py`, the ` maybe cleanup()` function uses an unlocked dictionary for cleanup throttling, which can lead to simultaneous ` cleanup tenant()` operations for a single tenant. The cleanup process reads and rewrites `transcripts.jsonl` without file locking or atomic replacement, while `report buffer append async` appends data without synchronization. In `main.py`, concurrent ` processing worker` tasks make this reachable, especially when `processing.worker count` is two or more and public-pool servers share a `server key`. When `report buffer.enabled` is set to true, these racing writes can cause loss, duplication, or corruption of transcript data and bypass `report buffer.retention seconds` pruning. This can result in unbounded disk growth, potentially exhausting storage and crashing the host or other shared services.