debug: improve vllm-direct endpoint to test context overflow

debug: add /api/debug/vllm-direct endpoint

fix: remove vLLM probe + move warmup post-planner

fix(warmup): fire vLLM warmup before planner so RunPod loads during planner+stones

fix(sse): send keepalive comments to prevent proxy idle timeout

deploy(l4): self-contained Riprap mirror