Run a reproducible NVIDIA vLLM serving benchmark

[DaBestCode]

Produce the serving evidence required for an official performance launch after the vLLM integration lands.

Acceptance criteria:

• Compare dense vLLM, RandKV, and at least one scored-selector baseline at matched model, dtype, cache budget, request trace, and generation parameters.
• Use a pinned container or lockfile and record GPU model/count, driver, CUDA, vLLM, PyTorch, model revision, and all command lines.
• Include warm-up and repeated trials; synchronize timing where applicable.
• Report request throughput, output-token throughput, TTFT, inter-token latency, peak GPU memory, and failure/OOM counts.
• Publish every raw trial in machine-readable form plus the aggregation script.
• Pair performance results with matched-budget quality results; do not optimize one by changing the other protocol.
• State negative or inconclusive results without filtering them.

This issue depends on the vLLM integration design and implementation. Comment with available hardware before claiming it.