fix(evaluation): record the cpu model in every run and qualify dense reproducibility

[poppycoderr]

Background

The reproducibility note said that dense rankings can differ between arm64 and x86_64, and it treated the hosted x86_64 runner as one reference machine. A CI run on dataset v2 then differed from the published m2-labelled-dataset report in 2 of 432 ranked lists. In each of the two, a pair of adjacent dense candidates with near-equal scores swapped places between ranks 7 and 10. No metric changed, and every sparse, hybrid and BM25 list was identical.

Hosted runners do not all have the same CPU, and run.json did not say which one a run used.

Change

 run.json
   "platform": "Linux-6.17.0-…-x86_64-with-glibc2.39",
+  "cpu": "<model name from /proc/cpuinfo>",

• ga-eval run records the CPU model. It reads /proc/cpuinfo where that exists and falls back to what Python reports.
• benchmarks/README.md and the evaluation strategy now state the claim that the evidence supports: published numbers are reproducible to the reported precision, and dense result lists are reproducible up to the order of near-ties.

The earlier wording was too strong, and this PR corrects it. No published number changes.

---

背景

可复现性说明里原来写的是：dense 排名在 arm64 和 x86_64 之间可能不同，并把托管的 x86_64 runner 当作同一台参考机器。后来有一次数据集 v2 上的 CI 运行，和已发布的 m2-labelled-dataset 报告相比，432 个排序列表里有 2 个不同。这两个列表里，各有一对得分几乎相同的相邻 dense 候选在第 7 到第 10 名之间互换了位置。没有任何指标变化，sparse、hybrid 和 BM25 的列表也全部一致。

托管 runner 的 CPU 并不都相同，而 run.json 此前没有记录一次运行用的是哪种 CPU。

改动

run.json 的变化见英文部分的 diff。

• ga-eval run 现在会记录 CPU 型号：有 /proc/cpuinfo 时从那里读取，否则用 Python 报告的值。
• benchmarks/README.md 和评测策略文档改成了证据能支持的说法：已发布的数字在报告的精度内可复现；dense 的结果列表可复现，但得分接近的候选之间的顺序除外。

之前的表述过强，本 PR 做了更正。已发布的数字没有任何变化。