Document the cd model on the package and the website (PR C of 4)

[cmg777]

Summary

Stacked on #3. It documents the cd model on the package and the website.

New pages

• docs/models/cd.qmd, rewritten in the shape of the sar page:

  • setup;
  • the spillover model: Examples 1–3, custom structures, several treated units;
  • estimation, including the efficient estimator;
  • inference;
  • structure checks: κ_A, selection, sensitivity;
  • an accessor table;
  • where scspill departs from the authors' application code and from mlsynth.

  Citations are mapped to the paper's v3 sections.

• Case study prop99-spillovers.qmd (executed and frozen, with a Colab notebook notebooks/prop99_spillovers.ipynb):

  • cd on Cao & Dowd's 51-unit panel: the direct effect and Nevada's leakage;
  • the year-by-year joint "no spillover" test;
  • the κ_A test, which rejects the paper's 13-state structure after 1989 and is reported as a warning;
  • structure comparison, the sensitivity analysis, and the 41-unit specification;
  • a side-by-side comparison with sar, including why the two models disagree on the sign of Nevada's spillover.

• Article articles/cd-replication.qmd (executed and frozen) reproduces, live:

  • v3's Section 6.2 statements: insignificant 1989–1990 effects, and Figure 3's break-even points 17.07 and 9.66;
  • the authors' frozen application output, with every P value equal;
  • their simulation statistics;
  • mlsynth.

Site and package updates

• Navbar: "Sudan case study" becomes a "Case studies" menu with Prop 99 and Sudan; the Articles menu gains the replication article. sudan.html keeps its URL.
• Pages: the model catalogue (with "Reach for it when / Look elsewhere when"), the landing page, Get started, the datasets article, For AI/LLMs, llms.txt, and the quartodoc section descriptions.
• Package files: README, the CITATION.cff abstract, package docstrings, and the cao_dowd fixture README, which now cites v3 and describes its checks as checks against the authors' code.
• Benchmarks: california_cd and cd_vs_mlsynth, registered in benchmarks/registry.py; both pass. The full REPORT.md is regenerating locally and will follow as a commit.

Validation

• make docs renders without errors.
• make lint typecheck test: 306 passed, 3 skipped.

🤖 Generated with Claude Code