add ornith-1.0-35b + titus-cyber-35b rows

[agent:cc] benchmarks: null borrowed conc=1 prefill/power on conc=4 rows (honest structured fields, not prose-only)

[agent:cc] benchmarks: refine concurrent-row notes (sustained-decode/one-shot caveat, adversarial review)

[agent:cc] benchmarks: 35B-A3B vLLM-NVFP4-EP vs llama.cpp-Q4 single-stream + 4-way concurrent (2026-06-22)

refresh: current model roster - 103 rows / 21 models from llm-lab card-bench 2026-06-21 (adds Qwen3.6-27B quant ladder + 35B sweep + candidates; smoke/trial excluded)

Add RLM long-context artifact

Publish RTX PRO 4000 benchmark scaffold