[koshikawa-masato]

This adds real-hardware measurements for two Tesla V100 PCIe 32 GB cards at 150 W each, comparing XeStrata xe0.1.40.2.1 with upstream Strata 0.1.41. It is a benchmark/diagnostic contribution, not a performance fix.

On this Ryzen 9 9950X / 128 GB system, XeStrata automatic placement selected 35+13 layers and gave 62.03 tok/s Japanese / 92.63 tok/s code, versus upstream's 74.10 / 107.87. Fixed 24+24 with STRATA_SM70_TABLE=1 improved this to 65.10 / 100.10. Automatic placement improved the 1,551-token prefill result from 699.70 to 766.77 tok/s.

Included:

• Japanese and English reports with hardware, commits, settings and limitations.
• All 36 individual measurements, CSV means, and exact synthetic request bodies.
• Selected startup/cache logs: asymmetric PCIe links (13.1 / 3.2 GB/s), placement decisions, resident experts and cache hits.

Important limits:

• contrib-llvm source build with CUDA 12.8 and sm_70; not the released deb.
• Two configured slots, one active request; 256 output tokens for Japanese/code and 64 for the summary.
• 8K/16K prefill and per-kernel profiling remain unverified. The “no XMX” log alone is not evidence of CPU fallback.
• The engines differ in version and output; this does not isolate a single cause or establish peak throughput.

Validation: recomputed all CSV averages from 36 rows; all recorded reused-token counts are zero. Markdown, offline links, codespell, gitleaks and whitespace checks passed. The repository lint runner's full REUSE check reports pre-existing missing notices in docs/BATCHING.ja.md, docs/ORCA.ja.md, and docs/XE.ja.md; no newly added file is listed. No runtime code changed, and production was restored after the tests.

[MistVVK]

ベンチ機能作成中のため、ベンチのプルリクは一旦お待ちください。