Inference-first 0.3.0: likelihoods as programs, optimizer mode, Pareto posterior, prior-owned encodings

[alexnodeland]

Closes the three caveats left by the 0.2.0 refactor (#19), completing the identity: fugue-evo is an implementation of fugue for running evolutionary algorithms as Bayesian inference (plus a standalone classic EC toolkit). Follow-up to #18.

Fix A — likelihoods as programs

GenomeLikelihood<G> replaces "fitness" as the conditioning abstraction: an observation program p(data|g) that may contain per-datum observe statements, factors, and — the capability the scalar factor could never express — latent nuisance parameters jointly inferred with the genome. The black-box mode survives as the FactorFitness adapter, now explicitly labeled as the Gibbs/generalized-Bayes posterior. EvolutionModel::new(prior, fitness) is source-compatible; from_likelihood accepts any observation program.

Demonstrated where it matters: GaussianRegression with NoiseSpec::Infer puts the observation noise σ at a latent site inside the likelihood — test_symreg_infers_noise_jointly generates data with σ=0.3 and recovers it from the particle traces while simultaneously finding the generating program.

Fix B1 — optimizer mode

EvolutionSMC::anneal(β_max, steps) continues the tempering ladder past β=1, built entirely from fugue's exported primitives (incremental reweight → normalize → resample → π_β-invariant rejuvenation → optional crossover sweep). On the sphere benchmark it reaches within ~1e-4 of the optimum while still reporting population uncertainty and log-evidence (examples/optimize_by_inference.rs, head-to-head with SimpleGA). Single-objective optimization is no longer the classic layer's monopoly.

Fix B2 — multi-objective as inference

ParetoScalarization puts the scalarization weight inside the model (uniform-simplex stick-breaking Beta sites): the joint posterior's marginal over genomes traces the Pareto front, and particle_weights reads each particle's front position off its trace — a posterior over the front, with evidence, which NSGA-II cannot express. Validated analytically: biobjective f1=x², f2=(x−2)² with Pareto set exactly [0,2] — >90% of mass on the set, both ends covered, and particles sit near their own weight's scalarized optimum x*(w)=2(1−w). (Weighted-sum = convex fronts; Chebyshev is noted future work, same architecture. CMA-ES deliberately stays a baseline — it is not a posterior sampler.)

Fix C — prior-owned encodings + cost

GenomePrior::trace_of encodes a genome under the prior's address scheme (grammar prior overrides with the exact inverse of its generative walk — pinned by an inversion test and a hand-computed PCFG score). EvolutionModel::score/to_weighted_trace now work for grammar trees, and EvolutionChain::init_from warm-starts a chain from any in-support genome — including a classic GA/GP result. MemoizedFitness (exact-key, shared-cache) removes repeated evaluations under replay-heavy inference.

Fix B3 — identity

• New classic feature (default on) gates algorithms/operators/population/hyperparameter/interactive/checkpoint/diagnostics/termination. MultiObjectiveFitness/ClosureMultiObjective moved to core fitness::multi_objective (re-exported from nsga2 for compat).
• --no-default-features --features std,ppl builds the inference layer with no classic EC code; std,parallel,checkpoint,classic builds classic with no fugue. Both configs check clean with --all-targets.
• Crate description and README now lead with the inference identity.

Validation

• Default: 9/9 suites, 0 failures (49 inference-layer tests incl. the new noise-inference, anneal-concentration, Pareto-front, encoding-inversion, and memoization anchors).
• Both minimal feature configs: zero errors, --all-targets.
• clippy 0 warnings, fmt clean, wasm crate unchanged and green.
• All prior anchors still green: EV-16 (+analytic evidence), EV-52, EV-90, EV-53, dead-chain, symreg recovery.

Version 0.2.0 → 0.3.0 (pre-1.0 breaking-minor: EvolutionModel generics; feature gating).

🤖 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.