#2216 bench: community report, one Arc A770 16 GB (xe) - main on the card and the A770 pull requests, same-day A/B

本文

## Title
Issue: #1545, #1606, #692 (Arc A770)

## Summary
A community benchmark report for one Arc A770 16 GB on the `xe` driver (Alchemist on `xe` is a gap in `docs/INTEL.md`: its Alchemist results are an A750 on `i915`). It records what `main` (`a0116b27`) does on this card and the A770 pull requests, each against a same-day build without it, plus the changes measured on this card and not taken. Docs only.

## What changed
- `bench/results/2026-10-13-community-arc-a770/README.md` (new).
- `docs/COMMUNITY_BENCHMARKS.md`: one entry in the community list.

## Extra Notes
The pull requests it refers to: #2190, #2193, #2203 (stacked on #2193), #2207, #2199, #2215; the hang: #2194. Two branches it mentions are pushed to the fork but not opened as pull requests, because they did not measure faster on `main` (spill-free kernels) or the problem did not reproduce on `main` (a lock on the prefill ring's release events).

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