v0.11.0

The quadratic model runs on the default simplex. The least-squares gradient is a closed form on any centred tight frame, so use_quadratic_model and trust_region no longer need the orthoplex's antithetic pairs: k+1 vertices per step instead of 2k. At a fixed 300K-evaluation budget over 6 seeds it reduces loss 0.178 against 0.079 plain and 0.156 for the orthoplex. It scores predicted against actual improvement, so it engages on a stationary objective at amortize_steps=1; the constructor now warns where it cannot.

Breaking. LowRankSubspace, LinearSubspace and FactorSpec are gone; use HybridSubspace or FactoredSubspace. Also removed: absorb_every, polystep.benchmarks.baselines with its evotorch dependency, and SinkhornSolver.solve(init_eps=). CMAAdaptiveSubspace now subclasses AdaptiveSubspace, so .base is gone.

Fixed. A zero transport row moved the particle to the coordinate origin, -inf costs tied with a real minimum, polytope_type='cube' at particle_dim=2 was read as an orthoplex by vertex count, and blockwise reported convergence on an all-NaN sweep.

Full list in CHANGELOG.md.