Add Image Generation & Diffusion track

[YashJayswal24]

Adds an Image Generation & Diffusion track: 3 chapters, 10 pages, 9 live simulators. Placed before the specdec track, because specdec/03-parallel-drafting already leaned on diffusion for DFlash and assumed the reader knew what it was.

Chapter	Pages	Simulators
1 — What Generation Asks For	the task · GANs · why they broke	ManifoldSlice, GANDuel, ModeCoverage
2 — Destroy It, Then Undo It	hook · forward process · noise prediction · one big step	MixingReversibility, ForwardNoise, ThreeTargets, StepBudget
3 — How Stable Diffusion Works	latent diffusion · conditioning · guidance	LatentCompress, GuidanceDial

New rendering path

lib/PixelCanvas.tsx is the repo's first canvas renderer. The site was 100% SVG/Plotly, which cannot carry a 24×24 frame (576 cells) across several panels at frame rate — that spends the whole budget in reconciliation. It renders grayscale in both themes on purpose: theme-inverting an image makes one sprite read as two different images, so only the frame is theme-aware.

Also lib/{diffusionMath,sprites,spectrum,pca,tinynn}.ts — all pure, no React imports, so they run directly under node --experimental-strip-types.

Every numeric claim was measured before the caption was written

That is the reason for the extra lib/ modules, and it repeatedly overturned the plan:

• The step-count sweep converges at T=4, not 1000. Distance-to-nearest-real runs 0.50 / 0.063 / 0.0032 at T = 1 / 2 / 4 and then sits at 0.0032 through T=200. Growing the training set 8 → 1024 changed nothing; injecting per-call denoiser error changed nothing. The page now says so outright, and explains that real systems need more steps because their denoiser is a learned approximation — with DDIM (1000→50) and consistency models (→1–4) as the evidence.
• A stronger GAN discriminator helps, rather than causing collapse: 6–8 modes and 9–15% off-data at 4 D-steps, versus 3–8 modes and 11–79% at 1 step. The folk story is about the 2014 minimax loss. What is reproducible is seed variance — 3 modes/79% off-data to 8 modes/11% across 8 seeds at one fixed config, with the loss curves looking the same either way. That became the chapter's thesis.
• Reverse KL does not mode-seek from an arbitrary start. Over seven starting points, forward KL lands at μ≈0.1, σ=2.65 covering 3/3 every time, while JS and reverse KL land on whichever mode they began beside. The start position became the widget's only knob.
• Guidance saturation artifacts do not reproduce with an exact score — distance-to-real holds at 0.003 from w=1 through w=15, while diversity does decay (13 → 8 of 16 images). The burned-out look in real systems is the learned score's approximation error, amplified.

Two controls that turned out to do nothing — a kernel-width knob and a dead Run button — were removed rather than shipped, with code comments explaining why so they aren't re-added. MILESTONES.md records all of it.

Second commit: two fixes

• Upstream's ch03 restructure removed mean-field-trap.mdx, so the two diffusion pages citing it now point at dspark/#the-trap-a-product-of-marginals (anchor verified in the built HTML).
• The redirect added for the old URL 404s on Pages. Astro emits the target verbatim into the meta-refresh, so /specdec/03-parallel-drafting/dspark/ without the /stable-learning base resolves to the domain root — the same trap as c005587. One-line fix; happy to drop it into its own PR if you'd rather.

Verification

• npm run build clean, 58 pages
• Whole-site link check: 3660/3660 internal links resolve, including the redirect
• Both themes checked by headless screenshot; canvas widgets are the first here that don't inherit theme automatically
• Library-level claims re-run against the actual modules, not eyeballed

🤖 Generated with Claude Code

https://claude.ai/code/session_01TCNCrkx7GQ85MV1A6b6bgU