#2709 STRATA_SM70_TABLE=1: real-world V100 result, +4.0% decode, no output change

Description

Tested `STRATA_SM70_TABLE=1` on Strata 0.1.41, Tesla V100 16GB, Qwen3.8-Flash-Next IQ3_XXS, n_ctx 196608.

Same machine, same probe, forward/reverse order:
- SM70 OFF: median **79.7** tok/s, best 81.4
- SM70 ON: median **82.9** tok/s, best 84.2
- SM70 ON (final confirm): median **82.5** tok/s, best 83.0

That's **+4.0%** decode, reproduced in both directions. For reference, 0.1.40.1 baseline was 81.5, so 0.1.41 + SM70 (82.5–82.9) is now above baseline.

Tool calls work normally, and quoted tool-call text inside thinking is still correctly ignored. I observed no output difference, and I'm keeping it enabled in production (VRAM headroom 748 MiB).

Note: I'm reporting observed behavior, not claiming bitwise-identical output — the bitwise claim is from the source comment, my side only verified tool-call behavior and daily output.