Chebyshev scalarization for non-convex fronts + evolution-as-inference explorable (0.3.1)

[alexnodeland]

The two follow-ups noted at the close of #18.

Chebyshev scalarization (inference::pareto::ChebyshevScalarization)

The weighted-max norm max_i w_i·(f_i − z_i) over an ideal point z, with either a latent weight (like ParetoScalarization) or a fixed weight (with_weight) for uniform front sweeps. Pinned at the theorem level on the concave front f1 = x, f2 = 1 − x² (Pareto set exactly [0,1], where every weighted-sum interior stationary point is a maximum):

• fixed-w weighted sum puts < 0.1 mass in the front interior (endpoint collapse, as the theorem demands);
• fixed-w Chebyshev concentrates on the interior front point x* = (√5−1)/2 ≈ 0.618 (posterior mean within ±0.08);
• a three-weight sweep traces low/mid/high front regions.

Working the test surfaced an honesty fix worth calling out: in the latent-weight models the w-marginal is tilted by exp(−s·m(w)) (the scalarized optimum's value at w), so annealing concentrates mass on the best-scoring front regions rather than covering the front uniformly — the conditional x | w is what tracks the front. Both scalarizations' docs now state this explicitly, and a new regression pins the conditional property for latent-w Chebyshev. (My earlier "the posterior traces the front" phrasing was stronger than the convex-case test had actually verified.)

"Evolution as inference" explorable (evo.fugue.run)

ExploreSmcInference (fugue-evo-wasm, now with the ppl feature + a direct fugue-ppl dep) steps the crate's real inference layer one tempering rung per step(): GaussianPrior program, twin-peaks Boltzmann target via FactorFitness, and fugue's SMC primitives (incremental reweight, ESS-triggered systematic resampling, typed-MH rejuvenation, crossover kernel). density_grid returns the exact tempered target so the heat overlay is the analytic density the particles should (and visibly do) match at every β; β MAX extends past 1 into annealed-optimizer territory; log-evidence accumulates live.

• Engine pinned by: determinism, ladder shape, analytic posterior mean and analytic log-evidence vs 400² grid quadrature, and annealing concentration (mean distance-to-nearest-mode — the honest metric for a multimodal target).
• Widget (docs/viz/inference.js) follows the site conventions exactly: wasm gate + no-fallback notice, lazy IntersectionObserver init, seeded scrub, controls → canvas → instruction → readouts, reduced-motion synchronous batch, theme-aware heat rebuild.
• Mounted on Architecture → Evolution as Inference; book.toml registers the viz file; no CI change needed (docs.yml already builds and stages pkg/).
• Browser-verified against the built wasm: autoplay runs the ladder to β = 1 with both modes populated and plausible ESS/log Z/swap readouts; Replay/Reset/Step all work (screenshots in session).

Validation

9/9 default suites + 5/5 wasm-crate suites green; both minimal feature configs check clean; my files clippy-clean and rustdoc-clean (the pre-existing --all-features workspace-clippy failure and 3 old wasm-crate warnings on main are untouched); fmt clean. Version 0.3.0 → 0.3.1 (additive).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HmWLBR9Zq6hPG5o7URDuNJ

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.