[Bench][SM70] Repair #375 without disabling QPN8

[yangzhuxinyzx]

Purpose

Repair #375 on the latest audited main without disabling the proven Qwen3.8 online-QPN8 performance route before matched model-level quality evidence exists.

Result

• Latest main, including [KV Cache] Zero compressed attention storage pages #373 and [Core][SM70] Repair Qwen4Exp PLE integration from #338 #374, is the source base.
• QPN8 remains default-on under the project performance policy.
• The safe [Bench][SM70] Record Qwen3.8 QPN8 quality/performance A/B #375 benchmark changes are retained:
  • repeatable engine argument overrides for matched A/B runs;
  • boxed-answer-first GSM8K parsing;
  • signed/comma-separated integer support;
  • rejection of non-integral decimal answers.
• The default-off runtime, sidecar, test, warning, and documentation changes are not carried forward.

Validation

• 27 focused benchmark/QPN8 tests passed; one CUDA extension test was deselected because this source checkout has no compiled vllm._C QPN8 operator.
• Changed-file pre-commit passed, including Ruff, formatting, mypy, SPDX, forbidden-import, and configuration gates.
• All commits are DCO signed.

[yangzhuxinyzx]

审计完成：锁定 head bb22d57；27 个聚焦测试通过（另 1 个仅因源码 checkout 无已编译扩展而 deselect），完整远端 pre-commit 通过。该修复保留 main 已验证的 QPN8 默认开启策略，只纳入安全的 GSM8K 参数覆盖和答案解析增强；源码树已核对为最新 main 加这两项 benchmark 改动。现在合并回 #375 分支。