queueTickets

Tracks every request that is waiting for or currently holding the inference semaphore. Streaming threads register before acquire; retry-based clients (sync LLAMA, Tesseract) register on the first failed tryAcquire. Tickets are removed when inference completes (in the finally block after release). The map value is the creation timestamp for waiting tickets, or Long.MAX_VALUE for running inferences (immune to stale cleanup).