v0.9.0

Added

• SparseDeltaEvaluator: a full-space candidate replaces particle_dim scalars, so its
  output differs from the base in at most that many columns. Propagates the delta through
  elementwise layers and goes dense one Linear past the perturbed one, instead of a
  dense bmm over N weight sets. Automatic for nn.Sequential of Linear and
  elementwise layers, with a per-chunk fallback.
• FactoredSubspace: perturbs each 2D parameter by dW = A @ B with B fixed, so
  x (W + A B)^T = x W^T + (x B^T) A^T and no candidate weight is built. Not the
  default: the fixed B confines perturbations to rank input directions.
• step(..., objective_token=...) marks which objective a step measures, so cost rows
  cannot be reused across minibatches. release_evaluator(), resync_from_model(),
  MinCostGreedySolver and TopKMeanSolver are public.
• MANIFEST.in, so the sdist ships a runnable test suite. CI runs it from the tarball.

Fixed

• api.train() never called register_evaluator, the only assignment of
  _cost_evaluator, _fused_inputs, _factored_evaluator and _sparse_delta_evaluator,
  so every fast evaluation path was unreachable through the training API.
• Cost-row reuse under adaptive_probes was per particle, but a candidate is the whole
  configuration with one row replaced, so a stagnant particle's row was still measured
  against every other particle's position. Reuse is now all or nothing.
• multifidelity_screen averaged direction contrast across particles, which carry
  independent rotations, so the kept set was chosen at random. Ranked per particle now.
• Tied weights raised on both evaluation paths. batch_unflatten emits the canonical key
  only; unflatten still carries aliases for load_state_dict.
• restore_best keyed on the OT cost when no callback was registered, restoring the step
  with the cheapest probe cloud rather than the best weights. It also snapshots to CPU.
• Chunk size ignored the weight set a non-fused subspace builds per candidate and the vmap
  activations. A 120K-param step peaked at 1661 MB in full space and 10565 MB in subspace
  mode; now 946 MB and 1461 MB.
• AdaptiveSubspace and per-layer rotation replaced the projection while the coordinates
  were non-zero, moving the weights with nothing evaluated behind them.
• CMAAdaptiveSubspace never rotated or absorbed, left its hyperparameters at 0.0 so
  the covariance update was inert, normalized the evolution paths by sigma instead of
  step_radius so C_diag decayed to its floor, carried state through basis rotations
  untransformed, and accumulated at bf16 where 1 + c_1*x rounds back to 1.
• Anderson acceleration subtracted dX instead of dX + dR, so anderson_depth > 0
  converged slower than no acceleration.
• num_particles = 1 froze the optimizer: balanced OT with one row forces a uniform plan.
  The default now selects SoftmaxSolver and warns.
• scale_cost divided before recentering, so adding a constant to every cost changed the
  plan. SinkhornSolver also reported ent_reg_cost in the wrong frame; the plan and
  duals were unaffected. The greedy and top-k solvers had the same reporting bug.
• Softmax paths recentred by the global minimum, so a row above it underflowed to -inf
  and returned NaN; the fused kernel did not recenter at all.
• sanitize_cost mapped -inf to the worst finite value instead of the best, and used an
  absolute penalty that entered the scale_cost='mean' reduction.
• state_dict captured only SolverState, so resume was not exact for amortize_steps,
  adaptive_probes, trust_region or HybridSubspace. Format is 3; format 2 loads with
  a warning.
• A stagnation absorb never cleared stagnation_count, redrawing the basis every step on
  a plateau. resync_from_model() left momentum and rotation state on the old anchor.
• Probe losses accumulated in FP32 unconditionally, so a float64 objective whose costs
  differ below FP32 resolution produced a constant cost matrix.
• Block-wise biased rotation raised on bf16 under mixed_precision=True. A zero descent
  direction produced a singular rotation collapsing +e0 and -e0.
• Smaller: the in-place evaluator dropped unknown parameter keys silently; the
  trust-region ratio scored a Newton step that was never applied; displacement history
  mixed bases; randomized PCA drew from the global RNG; rank transitions dropped
  HybridSubspace settings; data_dependent_init had the wrong sign; a float64 model
  with a subspace raised; a frozen model gave an opaque error; pyproject.toml declared
  torch>=2.4 where 0.2.0 announced torch>=2.8.

Changed

• adaptive_probes and adaptive_num_probe default on under block_strategy='monolithic'.
• use_csa is refused with a warning: the OT step is a deterministic descent direction, so
  CSA reads agreement as the normal state and sigma grows without bound.
• multifidelity_screen cuts the forward budget instead of only reweighting, and runs only
  when screen_fidelity/num_probe + screen_keep_ratio < 1. Needs a cheap closure from
  screen_closure_from(); api.train supplies one.
• Block-wise warns for options it ignores; grouped warns when combined with a subspace.
  LinearSubspace amplifies rather than dilutes the perturbation, as its docstring said.
• _step_blockwise claimed blockwise cuts OT cost from O(N^2) to O(N^2/L). The forward
  count is identical; L solves over P/L rows replace one over P.
• Removed get_sobol_rotation_matrices, HybridSubspace.reconstruct_base and
  reconstruct_batch_delta: no callers, no coverage.

Performance

• thin_qr pins basis construction to one thread: HybridSubspace setup for the MNIST
  example drops from 27.0 s to 97.8 ms. apply_biased_rotation uses modified Gram-Schmidt
  instead of torch.linalg.qr, 7-33x faster at these sizes. Compiled kernels take radius
  and epsilon as 0-d tensors, so a scheduled radius no longer recompiles every step.
• Example 06 on an RTX 5090: 1024 s to 845 s, shift-recovery +9.3 pp to +10.7 pp.

Tests, examples and docs

• 967 tests in 262 s with an intermittent timeout failure, now 941 in 8 s with none.
  tests/test_subspace_contract.py replaces per-class copies of the properties every
  subspace shares. Removed tests that could not fail. Suite verified by mutation: eight
  semantic changes to the source are all caught.
• tests/test_nondiff_data.py guarded nine tests on the pysat import but not on the
  experiments/ tree they import from, so they failed rather than skipped in the sdist
  when pysat happened to be installed.
• Examples pin CPU threads, check optional imports before use instead of failing after the
  run, and register the evaluator in hand-rolled loops. python-sat moved into the
  examples extra, which example 04 needs.
• release.yml lint and test commands had drifted from ci.yml and it lacked the CPU
  torch index; both now match. README and the docs corrected where they had drifted.