[12 points] Focused agent-weakness study for issue #455

[JunkaiWang-TheoPhy]

中文说明

改了什么

• 增加 issue [12 points] Paper study: identify agent weaknesses and measurement gaps / Agent 弱点研究 #455 的公开 benchmark-family evidence YAML。
• 增加 coding guide、blinded sample 和 independent review。
• 增加确定性分析脚本，输出 fine/coarse recurrence、measurement gaps 和 agreement。
• 在 next-draft technical report 中加入简洁的 agent-weakness subsection。
• 重新生成 next-draft PDF；不修改冻结 v0.9.0 PDF。

为什么改

这个 PR 将 issue #455 的研究证据、分析代码和报告内容放在一个聚焦的变更中，避免把历史 review 修复和其他 issue 的 report changes 混在一起。

重点审核

1. evidence YAML 的来源锚点与 benchmark-family 去重；
2. blinded secondary review 与 agreement 计算；
3. 报告中 8 个 demonstrated families、6/8 state-control、2/8 decision-execution 的表述；
4. next-draft PDF 是否与 YAML 和分析输出一致。

验证

• Ruff check / format check
• normalize-external、classify、build-data-release
• 完整测试：1293 passed
• focused tests：67 passed
• PDF 重新生成并检查

Closes #455

[ktwu01]

Nice evidence discipline: every row carries a source anchor, a counter-reading
and a counter-example location, and the 8+1+1 split reconciles.

The main thing this needs is a rebase, and it is not a defect in your work.
The branch edits scripts/build_technical_report.py and
scripts/build_system_evaluation.py, and neither file exists on current
main: the report has moved to the docs/technical-report/latex submodule.
The documented workflow here is also the superseded v0.9.0 one, where current
rules require frozen v0.11.0 inputs. The green CI on this PR ran against the
older base, so it does not tell us whether the study integrates with the
current report build. Rebasing onto main first will show what actually needs
rewiring.

Two smaller things:

1. data/agent_weakness_evidence.yml has radar_record_id: null on the
   swe_science_misguided_exploration and prbench_end_to_end_gap rows. Every
   other row links to a corpus record. swe_bench_science is already
   registered on main (data/model_cards.yml:1073), so the first should be
   linkable now without waiting on anything. Either link them or mark the gap
   explicitly so it reads as a known absence.
2. The null-ID validator does substring matching and skips validation entirely
   when the index is absent. docs/query-surfaces.md makes QueryService the
   single source of truth for lookups; routing the check through it would make
   a silent skip impossible.

Also please state the exact PDF regeneration command in the PR body; the
next-draft PDF lands as a binary diff.

---

Verdict: Comment. Review produced by an agent (Claude), reading the diff, the repo rule files and the cited sources, then re-checked against an independent adversarial pass. Posted as a comment rather than a blocking review.