Research-production instrument: node -> hypothesis -> code-verified r…

[ChatGPU]

…esult -> paper draft

Turns the atlas from a map into a research instrument, addressing the 'paper-production pipeline' intent:

• labs/research/harness.py: dependency-free (stdlib only, CPU, deterministic) harness that takes a falsifiable Hypothesis tied to a graph node, sweeps one independent variable, records a metric, decides whether the claim survived its own falsifier, and emits a results table + SVG figure + findings note.
• labs/research/exp_overestimation_bias.py: code-verifies the lens of insight:q_learning_max_is_optimistically_biased. Result reproduces the textbook effect — single-max bias grows 0.20 -> 0.83 as actions go 2 -> 64, while the double estimator stays ~0 — so the lens claim survives, by code.
• tools/scaffold_paper.py: given a node, assembles a paper-draft scaffold from the graph neighbourhood + research lens (research question, positioned related work, load-bearing assumption, gap, method primitives, cross-domain prior, experiment plan with a harness pointer, falsification criteria). Example drafts for UniAD, offline-RL-as-constrained-DP, differentiable-E2E, and conditional-generation planning in research/drafts/.
• Research lenses 56 -> 66 (completed RL/decision lens-wave fragment).

QC green: 0 dead links, 0/231 lint findings, math test passing, all 66 lens links resolve, seed validator OK. (Lands on the branch after the PR-23 merge.)

Pull Request

变更摘要 / Summary

（一两句话讲清这个 PR 的意图）

类型 / Type

• 新增 / 修改 spine 卡片
• 新增 / 修改 Tier-S/A/B 卡片
• 修改 docs/data/graph.json
• 修改 / 新增 lab notebook
• 修改交互站点（HTML/JS/CSS）
• 修文档 / Playbook / Concept Atlas
• CI / tooling

检查清单 / Checklist

• python tools/validate_graph.py 通过
• python tools/check_links.py 没有新出现的 404
• 所有改动到的 lab 在本地用 LLM_BACKEND=mock 跑过 nbconvert --execute
• 卡片新增 / 修改时同步更新了 concepts.md / playbook（如相关）
• 对外链接尽量是深度锚（PDF #page=N、YouTube &t=Xs 等）
• 采纳外部建议时已在相关卡片加入 contributors: [...]，或说明不需要标注

备注 / Notes

（截图、性能数字、与 AGENTS.md 的对照说明等）