Nemotron 3 Nano 4B bench records

[webdevtodayjason]

Nine schema-1 records from tonight's run, no code changes.

• Spark-4 (GB10) through AINode, vLLM 0.27.1 BF16: throughput, agentic (17/23), decide (94.5%).
• Pollux, one V100, llama.cpp with NVIDIA's Q4_K_M GGUF: throughput (153 tok/s), agentic (20/22), decide (91.8%).
• Pollux, one V100, onecat-vllm src-full FP16: throughput (84 tok/s single, 463 at 8). Its agentic record reads 1/22 because that build's nemotron_v3 parser files every answer as reasoning; it is kept as a record of that, not as a model score.
• Jeb 4B v2 decide baseline run alongside (97.3%). Its placement block may name Spark-4 though the proxy served it from another node.

The V100 runs were outside AINode because of #296. Write-up: https://artifacts.semfreak.dev/a/ainode/nemotron-nano-4b-fc67d959/

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}