Markov: sample-cloud overlay, About-the-math guide, and a Bayesian posterior target

[kkd16]

Follow-up to the Markov sampling studio (initial version merged in #664). All changes stay inside projects/markov-lab-q8f3/.

What's new

• Long-exposure sample cloud — an accumulation canvas splats every state the chain visits with additive blending, so the sampler is seen repainting the target density one dot at a time (toggle: cloud). Turn density off and watch it rebuild the distribution from scratch.
• Keyboard shortcuts — space run/pause, s step, r reset.
• "About the math" guide — a modal explaining all eight samplers, how to read every diagnostic (ESS, ESS/eval, split-R̂, autocorrelation time τ), and a guided set of experiments.
• A real Bayesian posterior target — a sixth target whose two coordinates are the intercept and slope of a logistic regression fit to 18 near-separable data points (numerically-stable log-likelihood + exact gradient, N(0, 3²) prior). Because it satisfies the same Target interface, every sampler and every diagnostic now performs genuine posterior inference for free: the running mean is the posterior-mean estimate and the 95% CI is a real credible interval.

Notes

• Verified locally with node scripts/verify-project.mjs markov-lab-q8f3 — scope + conformance + lint + build all green.
• Smoke-tested in headless Chromium: the cloud overlay, the About modal wiring, and the logistic posterior all render live (HMC exploring a broad, prior-regularised correlated ridge, exactly as expected for near-separable data).
• Branch was restarted from the latest main after Add Markov — a from-scratch Monte-Carlo sampling studio #664 auto-merged, so this PR contains only the follow-up diff.

🤖 Generated with Claude Code

---

Generated by Claude Code

[github-actions]

✅ Scope + lint + build passed.