feat(synthesis): support (Array Float) holes in FloatHoleNG

[alcides]

Stacked on #403 (base: claude/genomic-ng-float-holes), which is itself stacked on #401. Review/merge those first; this PR's diff is only the third commit.

Summary

Extends FloatHoleNG (#403) beyond scalar Float holes to holes of type (Array Float) — a single array hole becomes one optimization dimension per element, which is the most natural framing for vector-valued continuous optimization.

Each hole now maps to a contiguous slice of one flat float vector:

• a scalar Float hole → one dimension → substituted as a Float literal;
• an (Array Float) hole → its length k dimensions → substituted as a list-valued Array Float literal ([v₀ … v_{k-1}], which the backend evaluates directly to a Python list).

Determining the length

The length comes from a fixed-size refinement, {a:(Array Float) | size a == k} — the same Array_size measure the multi-objective decorators use. The extractor walks the refinement for an Array_size == int equality (either operand order, inside conjunctions). An array hole without a fixed-size refinement is rejected with a clear message (its length is unknowable).

Precondition

Now stated in float dimensions (≥ 2) rather than hole count, so a single (Array Float) hole of size ≥ 2 qualifies, while a lone scalar Float hole still routes to the grammar backends.

Benchmark — examples/synthesis/float_ng/sphere_array.ae

Sphere over a single size-5 (Array Float) hole (scripts/bench_float_ng.py, 3s, 2 seeds):

benchmark	ng_float (NGOpt)	ng_float_cma	gp
sphere_array	0	2.3e-18	—

Both float backends reach the optimum; gp cannot produce a usable array for this hole.

Test plan

• uv run pytest tests/float_ng_test.py — 15 passed, incl. _is_array_float, _array_length (size extraction / reversed operands / no-fixed-size), end-to-end array convergence, and the updated precondition message
• genomic_ng / knee_point / synth_fitness suites still green
• pre-commit (ruff + mypy) clean

🤖 Generated with Claude Code