#2006 bench: RTX 3080 Laptop 16 GiB, Flash-Next Q2_0 on Windows

Beschreibung

Adds the community benchmark measured on 2026-10-11 using tools/strata_bench.py, with README.md, per-request results.json and the startup engine log.

Hardware: RTX 3080 Laptop GPU (16 GiB VRAM), Ryzen 9 5900HX, 63.9 GiB RAM, Windows 11. Strata and engine 0.1.43 at 7695859a, Qwen3.8-Flash-Next Q2_0, 65,536-token context, one slot, unchanged installation configuration, thinking off and temperature 0.

Three runs per test after warm-up: short chat decode median 65.4 tok/s; fresh 4K/16K prompt throughput 878.6/1379.7 tok/s and TTFT 4.61/11.68 s. Three agent sessions: follow-up decode median 73.8 tok/s, conversation cache hit median 68.5%.

Validation: all three files read, report checked against JSON summaries, commit verified to add only those files. Engine code and defaults remain unchanged. No backend build is needed for report files.

Limitations: speed only, no correctness checks or same-day comparison with another build. No concurrency or 32K/long-context test. Storage, background workloads, power limits and CUDA version were not recorded. No performance improvement claim is made.