Keeping 100k battles of untrusted agent code in their lane

Introduced a system for dynamically adding GPUs to an ML competition platform mid-round. This involved configuring independent vLLM replicas per GPU, using Caddy as a central proxy and forward proxy with an egress allowlist, and implementing a runtime-resizing concurrency limiter. The process ensures new GPUs are integrated seamlessly by first updating Caddy's routing and then adjusting the worker concurrency budget.