#1432 bench: Arc Pro B65 Gen4 results on Strata v0.1.40.2

Beschreibung

Adds Gen4 x16 benchmark results for one Arc Pro B65 on the unmodified official v0.1.40.2 release (`e8ca9afd`). Original full 512-expert Flash-Next IQ2_XS, INT8 KV, protected pinned expert mirror, GPU-only execution and confidence-limited MTP4.

The 8K/512 profile measures 43.49 decode tok/s at 7K input; 262K/512 measures 42.38 on the same workload. At 8K, prefill 4096 reduces 7K TTFT from 22.23 to 8.47s, but changes long continuations and has a 16.2% decode regression in one coding cell. Normal serving retains prefill 512. v0.1.40.3 landed during testing and is not included in these pinned results.

Three fresh runs per profile in each of two suites: matched 512/7000-input, 640-output tasks and separate public 20/2185-input, 256-output fixtures. The report includes all repetitions, native prompt/decode timing, median/range, startup/cache state, exact hardware/build/artifact/configuration pins and portable reproduction scripts. Full-window capacity checks are reported separately.

Native reference, API, lifecycle and hardware checks passed for the measured profiles. The registered kernel suite has four classified failures (26/30 pass); the report explains those limits and the separate passing IQ/model-stream checks. No engine patches or tolerance changes.

Codex, an OpenAI AI agent, ran the tests and drafted and submitted this report with my approval. The measurements come from my B65 hardware.