#2675 bench: community report - RTX 3090, Core Ultra 9 285K, 95 GB, Windows 11, IQ3_S on v0.1.44.1

Description

## Title
Issue: none

## Summary
A community benchmark report: RTX 3090 24 GB, Core Ultra 9 285K, 95 GB RAM, Windows 11, Flash-Next IQ3_S at 262K context, on a v0.1.44.1 source build. Also, labelled as such, same-day A/Bs of engine settings and of two requests at once.

## What changed
- `bench/results/2026-10-15-community-rtx-3090-iq3-s-0.1.44.1/`: `README.md`, `results.json` and `engine-log-excerpt.txt` from `tools/strata_bench.py --long`, with the hardware fields the tool could not read filled in (build, CUDA, storage, background load).
- The README adds a section of A/Bs (prefill chunk, `--spec`, `--lookup-chain`, `--kv`, `--kv-resident`, up to 84K tokens) and of `"parallel": 2` with `--batch-mtp`. Those ran on a v0.1.43 build with local changes (hibernate and start-up patches, none in decode or prompt kernels), and the section says so.
- `docs/COMMUNITY_BENCHMARKS.md`: one entry.

## Extra Notes
The finding that may matter most for the docs: on this 24 GB card, `"parallel": 2` with `--batch-mtp` was slower than one request at a time in every case. Two ~100K-token conversations together gave 55.8 tok/s in total against 70.3, and a single request ran at 69.8 against 83.2 tok/s with the second slot on. The +31-39% reported for batch slots was measured on 48 GB cards.

Not tested, and stated in the report: contexts above 84K in the A/Bs; `"parallel": 2` without `--batch-mtp`, short prompts with two slots, more than two slots. Home folders show as `<home>`; there are no user, host or network names in the files.

🤖 Generated with [Claude Code](https://claude.com/claude-code)