#2867 bench: Intel Arc Pro B70 SYCL IQ3_XXS on 0.1.45 (LXC, 256K config)

Beschreibung

## Results-only community report

Single Intel Arc Pro B70 (nominal 32 GB), native SYCL Release AOT bmg-g31, Strata 0.1.45 at 477ac13c, original Flash-Next GSQ-RCO IQ3_XXS with a 262,144-token configured context.

- Official strata_bench.py, three runs per case, warm-up/loading excluded; fresh prompt cache reuse 0.
- Prefill medians: 1,349 / 1,446 / 1,407 tok/s at actual 4,019 / 16,058 / 32,049 tokens.
- Pooled short-chat decode 42.8 tok/s (36.3–55.9, nine capped requests); median TTFT 0.85s.
- Three agent sessions: follow-up decode median 51.8 tok/s, cache hit median 68.85%.
- Separate upstream needle checks: 6/6 at nominal 32K/128K and 10/50/90% depths; timing can include prefix-cache benefits.

Shared Proxmox/LXC host, not an isolated benchmark system: Ryzen 9 PRO 3900, host usable RAM 62.69 GiB, container limit 40 GiB / 16 logical CPUs, models on NVMe. README distinguishes container auto-detection from physical host facts and documents missing power/peak-memory measurements. No full-256K quality claim.

Includes per-request JSON, recall results, matching startup-only excerpt, public build/profile hashes, and reproduction instructions. Credentials, private hostnames/network details and private workload content are excluded. Seven new report files only; no engine or headline-performance changes.