mixle v0.7.0

Workstream: generic AI-capability platform pieces on top of the core estimation engine (task
decomposition, cross-modal reasoning, self-improvement loops, a system facade), plus a hardening
pass across the automatic-inference and design-of-experiments subsystems.

Added

• New model families: PINNRegression (physics-informed neural network leaf), HamiltonianNet
  (conservation-law-preserving dynamics), make_deep_set (permutation-equivariant networks),
  monotonic MLPs and input-convex energy networks, build_product_energy_net (energy-based product
  of experts), CopulaDistribution (arbitrary marginals + a Sklar dependence core),
  GatedMixtureDistribution (input-dependent mixture-of-experts weights).
• Task-decomposition and agent-facing workstreams: plan models fit as Markov chains over agent
  traces, an outcome-trained decomposer, a minimal orchestrator loop, tabular Q-learning and
  maximum-entropy inverse RL, a local model registry, ExecutionTrace with bit-identical replay, the
  Receipt object (ledger + trace + calibration + provenance, offline re-verifiable), and the
  System facade (answer/ingest/improve).
• Cross-modal reasoning: ModalityView, per-edge conditional-transport premise checks, belief walks
  across chains of verified transports, cycle-consistency as a self-supervised abstention signal,
  task-sufficient projection, information-gain retrieval, and the workstream-F flagship harness.
• Self-improvement / knowledge-accumulation loops: the collapse monitor shared across amplification
  loops, the composition operator, DesignModel's cross-round what-works prior, diagnosis-directed
  correction, a knowledge-accumulation flywheel measurement, degradation-policy handling for fault
  modes, and cost-aware routing threshold selection.
• doe: VerifiableOracle + the design-test-learn loop, noise-robust incumbent selection for
  Bayesian optimization, and a budgeted propose-verify-retrain loop over a discrete design space.

Fixed

• mixle.utils.automatic: crashes on empty/degenerate input (ZeroDivisionError in the Poisson/
  Gaussian/log-normal estimator builders), every distribution detector silently dropping its own
  already-computed fit and the caller's pseudo_count, an IndexError on always-empty sequence
  fields, a modality-fingerprint diagnostic that could contradict the actual estimator built, and the
  model-suggestion logic ignoring its own held-out validation signal when it disagreed with the
  in-sample BIC pick.
• mixle.data: Boolean.coerce silently inverting string-typed values (bool("False") == True),
  Schema.conform_record silently truncating mismatched records via an unchecked zip, an
  inconsistent tuple-vs-list adjacency coercion in the graph data source, and MaterializedSource
  silently accepting non-reiterable one-shot iterators.
• mixle.doe: an unguarded Cholesky decomposition crashing on singular/near-singular input
  covariance, a Morris-screening ZeroDivisionError on a degenerate grid, a silent-NaN Gaussian
  Process surrogate when fit with zero observations, an infinite loop given a zero-cost fidelity in
  multi-fidelity optimization, several proposal functions crashing instead of validating
  n_candidates, TuRBO overshooting its evaluation budget on trust-region restart, silent
  batch-truncation/duplication under an oversized batch request, and BayesianOptimizer.ask()
  re-dispensing duplicate initial-design points in async/parallel ask-before-tell campaigns.
• A circular import that broke mixle.inference entirely; a mixture-correction term error in
  explain()'s decision-margin ledger; a stale capacity.py embedding-head rung mismatch; CI
  flakiness in EM's log-likelihood computation and a de-flaked mixture-of-trees test; Python
  3.10-specific abstention timing in the oracle-timeout path.
• Release-verification pass (found via a fresh, non-editable venv install of the built wheel --
  never caught by the dev environment, which has every optional extra installed): import mixle
  was completely broken (a missing-import NameError at class-definition time in
  mixle.models.dpo_leaf cascaded through mixle.models.__init__'s eager import chain into nearly
  every module), plus the same missing-import/stale-duplicate-method pattern recurring across 7
  sibling model files and pinn.py; 8 further modules importing torch-gated names unconditionally
  at module level (undermining their own already-correct optional-torch guards); a real
  DPOAccumulator.value() bug (weights returned as a list, not an array); a data-shape bug in
  zero_shot_bootstrap's generic neural-density fallback (a 24-dim row was split into 24 scalar
  fields instead of one vector field); a stale test fixture double-wrapping Registry.tier_stack's
  frontier callable; a layering violation (mixle.experimental.long_context_eval importing upward
  from mixle.ppl); and 2 more test files with the same unguarded-torch-import bug. Also: numba's
  tbb dependency floor made pip install mixle[numba]/mixle[all] uninstallable on Apple Silicon
  (no arm64 wheels) -- now platform-gated; ray and lightning were used by real, documented
  optional backends with no corresponding pip install mixle[...] extra -- both added.

Changed

• pinn_leaf.py renamed to pinn.py; the "leaf" suffix dropped from PINN naming throughout.
• Several mixle/doe tests re-marked slow (heavy Monte Carlo / neural-density fits) so the default
  fast test gate stays fast; a real duplicate-training bug (an estimator refit once per test instead
  of once per class) fixed alongside the re-marking.