[Core][SM70] Keep Qwen3.8 QPN8 default-on after #375 A/B

[yangzhuxinyzx]

Purpose

Repair #375 on the latest audited main after its post-review ancestry merge
reintroduced the Qwen3.8 online-QPN8 default-off change.

Audit decision

• Keep online QPN8 default-on for the exact SM70/TP4/B1/no-MTP contract.
• Preserve an explicit VLLM_SM70_QWEN4_EXP_ONLINE_QPN8=0 diagnostic opt-out.
• Retain the safe repeatable engine-argument and boxed-answer benchmark changes.
• Record the matched 96-request A/B instead of discarding it:
  • GSM8K: 96/96 correct with QPN8 both off and on;
  • mean steady decode: 67.627 to 79.105 tok/s (+16.97%);
  • median steady decode: 67.640 to 79.369 tok/s (+17.34%);
  • aggregate repeated-4-gram mean: 0.0394 versus 0.0410;
  • one correct but substantially longer/repetitive output remains documented
    for broader long-output monitoring.
• Do not use token-hash identity or projection relative-L2 alone as a rejection
  criterion for this sampled inference optimization.
• Exclude the default-off runtime, sidecar, warning, and unit-test changes.

Validation

• 27 focused benchmark/online-QPN8 tests passed; the native CUDA-extension test
  was explicitly deselected because this clean source checkout has no compiled
  vllm._C operator.
• Changed-file pre-commit passed, including Ruff, formatting, markdownlint,
  mypy, SPDX, forbidden-import, and configuration gates.
• The repair was built from public main at
  0ca71115639bb4ddad7c7a818825c1261cd7433e and preserves [Bench][SM70] Record Qwen3.8 QPN8 quality/performance A/B #375 head
  5cb8dca64903f386371595d2a7064a8072c61de4 as ancestry without changing the
  audited source tree.
• All repair commits are DCO signed.

[yangzhuxinyzx]

审计完成：锁定 head 4d10309；27 个聚焦测试通过（原生扩展测试因干净源码 checkout 无 vllm._C 而明确 deselect），changed-file 与完整远端 pre-commit 均通过。源码树仅含 benchmark、测试和 A/B 文档，QPN8 运行时默认保持 main 的开启策略。现合并回 #375。