feat(synthesis): GenomicNG — Nevergrad grammatical-evolution backend

[alcides]

Summary

Adds GenomicNG, a synthesis backend that drives Aeon's grammar-guided synthesis with Nevergrad derivative-free optimizers.

A derivation is encoded as a fixed-length integer codon genome — exactly the representation geneticengine's GrammaticalEvolutionRepresentation already decodes. Nevergrad optimizes that integer vector; each genome is decoded into an Aeon Term by the existing GE mapper and scored by the existing validate + @minimize fitness. So the entire synthesis stack (grammar generation, phenotype decoding, fitness) is reused — only the search strategy changes.

Flags

• -s ng / -s genomic_ng — NGOpt (Nevergrad's recommended meta-optimizer; default)
• -s ng_cma, -s ng_de, -s ng_pso — CMA-ES, Differential Evolution, PSO

All four route through one class differing only by optimizer name, so adding more is a one-line case.

Behaviour

• Single- and multi-objective goals. Multi-objective resolves the Pareto archive by knee-point compromise selection (mirrors GESynthesizer).
• Robust to the ill-typed/un-evaluable candidates a black-box search constantly proposes: they are penalised, not fatal.
• Early-stops once problem.is_solved holds, instead of burning the rest of the budget.
• nevergrad is an optional synthesis-ng extra, imported lazily — the default install is unchanged.

Tests

tests/genomic_ng_test.py (15 tests): objective re-orientation, knee-point selection, factory routing for all 5 flags, and end-to-end single/multi-objective + alternative-optimizer synthesis.

Benchmark

scripts/bench_genomic_ng.py compares GenomicNG against GP. On a 10-example, validate-heavy suite GP wins decisively (8/10 solved vs 4/10) — the expected result, since grammar-guided tree search beats flat-vector black-box optimization for finding valid typed programs, especially on validate-only goals with no fitness gradient. GenomicNG's value is on problems with a strong continuous objective and high valid-candidate density.

Test plan

• uv run pytest tests/genomic_ng_test.py — 15 passed
• uvx ruff check clean; pre-commit (incl. mypy) passes
• End-to-end runs on examples/synthesis/{dummy,multiobjective}.ae with -s ng and -s ng_cma

🤖 Generated with Claude Code