#2748 Community benchmark: Apple M5 MacBook Air, 32 GB, Q2_0 (Metal, experts streamed)

Description

## Summary
Results-only submission: two runs of `tools/strata_bench.py` (default suite, 3 runs per test) on a MacBook Air (Apple M5) with 32 GB of unified memory, Flash-Next Q2_0, 32K context. MAC.md lists only the M5 Max (128 GB) and the M4 Mac mini (32 GB) as measured, so this adds a 32 GB M5 data point for the streaming path.

## What changed
- `bench/results/2026-10-16-community-macbook-air-m5-32gb-q2_0/`: first run, normal desktop load, server already running.
- `bench/results/2026-10-16-community-macbook-air-m5-32gb-q2_0-run2/`: second run after a macOS update and restart with apps closed; the benchmark started the server itself.

Each folder has the tool's `README.md` and `results.json`, plus a "Notes added by hand" section.

| | Run 1 | Run 2 |
| --- | ---: | ---: |
| Expert arena | 5.41 GB | 8.68 GB |
| Chat decode tok/s | 9.6 (8.8-10.2) | 17.7 (14.7-19.5) |
| Chat TTFT s | 3.57 | 2.17 |
| 4K prompt tok/s | 198.0 (187.2-201.7) | 165.0 (158.1-215.8) |
| 16K prompt tok/s | 182.1 (177.8-199.2) | 172.1 (166.7-174.1) |
| Agent decode tok/s | 8.9 | 9.7 |
| Agent cache hit | 68.7 % | 68.7 % |

## Setup
- Hardware: MacBook Air (Mac17,3), Apple M5 (10-core CPU), 32 GB, model on the internal SSD. On AC power, 23 °C ambient, no forced cooling. The Air is fanless, so an actively cooled M5 (e.g. a Mac mini) may sustain higher clocks over long runs.
- macOS 27.2: build 26B5101f (run 1), 26B5112a (run 2).
- Strata 0.1.44.1 at `296b59d2`, engine `0.1.44.1-metal`, built from source by `./setup.sh --yes --model Q2_0 --no-start`.
- Q2_0 with the experts' copy (`.experts`, 31.64 GiB), no MTP draft layer, prompt chunks of 1024.

## Extra Notes
- `strata_bench.py` reports "CPU only" on a Mac because it detects GPUs through `nvidia-smi`/`rocm-smi`. I corrected the title and GPU line of both READMEs; `results.json` is unchanged.
- Run 1 had only 12.9 GB of memory available at start, so the arena was 5.41 GB. That matches MAC.md's note about smaller arenas decoding slower.
- Prompt throughput was lower in run 2 despite the larger arena. No swap was used and macOS recorded no thermal or performance warnings, so I can't explain it. The 4K runs varied widely (215.8, 158.1, 165.0 tok/s).
- A background sync client was running during both runs. Speed suite only; no correctness or needle checks were run.

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