feat(api): exact Sobol' pair estimator + domain⊥distribution split (T51pq, T52xx) → 0.1.0a8

[JavierGOrdonnez]

What

The two backend tranches mmux_vite grew after the v1 port, ported to itis-sumo as one review train, per the owner decision to land both the T31rb estimator and the V26dd domain⊥distribution split now (no defer).

1. 0ef174d — exact arbitrary-d 2nd-order Sobol' pairs + order masses (T51pq, V48tr, B20qt)

The v1 port shipped mmux_vite's closed-form second-order identity, which mmux_vite B26nc proved underdetermined for d≥4 (pair sums collapse to 0 at d=4, negative beyond — typical 4–25-parameter studies produced invalid interactions) and B28pp proved non-translation-invariant (drift ∝ offset when mean≫std, e.g. stress in Pa). Ported from mmux_vite PR#649 (upstream/develop@6eda782d):

• _saltelli_abc — A/B/C blocks from one 3·d scrambled Sobol' stream (C feeds the pair designs)
• _saltelli_pair_designs — exact U^ij/V^ij mixed designs; surrogate cost n·(2 + d + d(d−1)), the documented price of pairs EXACT for any d
• _sobol_algebra — scipy saltelli_2010 parity pinned by test (pooled A/B mean/var, centered products → single estimator for all d; retires the d==1 branch and the runtime scipy.stats.sobol_indices call)
• _sobol_joint_bootstrap — ONE shared-row resample recomputes every index AND the M1/M2/R masses per replicate (estimator covariance + closure preserved)
• public: OrderMasses + SobolResult.order_contributions, null exactly at zero sample variance (undefined fractions, never fake (0,0,0))
• tests: the historic inlined-duplicate Ishigami math replaced by PRODUCTION-helper benchmarks (Ishigami, additive d=8, pair-interaction d=5, pair-quadratic d=10) with 3·CI-scaled tolerances + B26nc degeneracy regression

2. 59cf3c9 BREAKING — domain⊥distribution split (T52xx enforces V26dd)

evaluate_sobol sampled Saltelli A/B from modeller distributions — the conflation the SPEC scope line retires. Now the sensitivity box is domain vocabulary end to end:

• evaluate_sobol(samples, variables, response, *, domains=None, ...); distributions= is gone (it stays the exclusive input of evaluate_uncertainty)
• omitted boxes auto-inferred from observed sample bounds + detected scale; explicit boxes echoed verbatim; unknown names → SumoInputError; a column constant in the samples is pinned (never a fabricated box around it, V47st-consistent)
• draws are uniform / log-uniform (V44ls) via the shared scale_distribution; log⊗normal Sobol sampling and any mean±3σ-derived box die
• SobolResult: distributions → domains + fixed + effective_config (V21pf inspectable)
• flip matrix: Sobol's scale flip rides DOMAIN boxes; the NORMAL matrix has no Sobol row by design (2 distribution entry points remain, asserted)

Verification

• 387 passed full standalone suite (real itis-dakota 6.24.7, nothing mocked), incl. 26-test analytical tier and the 14-test test_sobol_indices.py gate
• algebra pinned vs scipy to 1e-9 + translation invariance at ~100× output offset
• V&V report: I7/I8 amended, I11 added, new "Estimator tier" section + M1/M2 gate rows
• SPEC: V48tr, V26dd enforced, V9gh amended, T51pq/T52xx, B20qt; PORTING.md branch inventory updated

Release

pyproject bumped 0.1.0a7 → 0.1.0a8 (valid alpha, newer than all tags). After merge to develop: tag v0.1.0a8 → CI auto-publishes to TestPyPI, which unblocks the mmux_vite Phase C rewire (pin itis-sumo==0.1.0a8, delete the inline dakota/ + data_preprocessor/ backend).

Consumer impact (mmux_vite Phase C/D)

The FE UQ-panel feeds constant|uniform|normal per Sobol variable; post-V26dd the Sobol section migrates to a bounds editor (T27fr's remaining item). Uniform → DomainSpec(min,max); constant/narrow → omit or pin; normal → drop (auto-inferred domain is the scientifically sane default the split exists to enforce).