Problem: The publisher dedicated one permanent thread per model. When the number of models exceeded the thread pool size, excess models were never processed (thread starvation).

Solution: Replaced the one-thread-per-model design with a shared thread pool and one-shot tasks:

• Discovery (every 30s) — detects new models in artery_model_infos, initializes them once; removes models that no longer exist.
• Polling (every 0.5s) — submits a finite process_model task for each idle model. The task drains all pending batches, then returns the thread to the pool.
• @busy_models set prevents double-submission; ensure guarantees cleanup on errors.

All models now get fair scheduling regardless of pool size. Busy models still process consecutive batches without latency penalty.