design: minimum-aberration generator selection — fractional factorials reach textbook resolution

[ChrisW09]

Fixes #10.

The greedy "largest subsets first" generator assignment created short defining words via XOR (e.g. ABCDE ⊕ ABCF = DEF), yielding resolution III for 2^(6-2), 2^(7-3), 2^(7-2), and the default 2^(8-4) — all of which have standard resolution-IV designs — despite the module docs promising resolution-maximizing generators.

Approach (the word-length pattern depends only on the set of generator subsets, so the search is over combinations(pool, p)):

1. Exhaustive minimum-aberration search when C(|pool|, p) · 2^p ≤ 600k: the XOR-closure of the generator words is computed with int bitmasks in gray-code order and candidates compared by word-length counts (lexicographically fewest short words — which maximizes resolution as a corollary). This covers every small-p case, i.e. all the designs where resolution > IV is achievable (2^(5-1)_V, 2^(6-1)_VI, 2^(8-2)_V, …).
2. All-odd-size subsets otherwise: every defining word (odd subset + one added letter, and all XORs) has even length ≥ 4 — the classical guaranteed resolution-IV construction.
3. Largest-first greedy only when a heavily saturated design exhausts the odd pool (resolution III is unavoidable there).

Verified against the standard maximum-resolution tables (Montgomery Table 8.14 / NIST) for every combination of k ≤ 15 and runs ∈ {8, 16, 32, 64} where a fraction exists — all match exactly, configs stay unique, and the slowest case takes ~30 ms:

k= 6 runs= 16  res=4 (max 4) | k= 7 runs= 16  res=4 (max 4) | k= 8 runs= 16  res=4 (max 4)
k= 5 runs= 16  res=5 (max 5) | k= 8 runs= 64  res=5 (max 5) | k= 7 runs= 64  res=7 (max 7)
k= 9 runs= 32  res=4 (max 4) | k=15 runs= 16  res=3 (max 3) | ...

The two existing tests (saturated k=7 → III, half-fraction k=5 → V) still pass; a new test pins six representative designs against the tables.

Test: full cafe-core suite — 109 passed (1 pre-existing failure = #4, fix in PR #34).

🤖 Generated with Claude Code