0.38.0

This release focuses on two things:

• Faster optimization loops: surrogates are now updated incrementally between iterations instead of
  being retrained from scratch, and the infill / hyperparameter optimizers move to pure Rust (basin).
• A cleaner Python API: standard Python exceptions, consistent names and shapes, per-constraint
  tolerances, and results you can feed back to the optimizer. Old names keep working with a
  DeprecationWarning and will be removed in the next release.

Highlights

Adaptive surrogate update (#454, #461, #462)

• Surrogate models are persisted across iterations. When the clustering is fixed and the new data stays
  consistent with the current model (z-score check), the GP is updated with the new points without
  re-optimizing its hyperparameters; otherwise a full training is triggered.
• TREGO local steps reuse and incrementally update the persisted models instead of retraining all
  surrogates from scratch at each local iteration.
• Redundant retraining at the start of each EGO step is skipped.
• The persistent feature is now always on in egobox-ego and egobox-moe.
• Fixes along the way: derived constraint models updated in the wrong space, stale virtual observations
  with q-batch, duplicated training data with mixed-integer variables.

Pure Rust optimizers by default (#458, #459, #460)

• basin optimizers (COBYLA, SLSQP) replace argmin and are used by
  default for infill optimization and GP hyperparameters training.
• The C-ported NLopt crates are still available behind the opt-in features c-cobyla and c-slsqp.
  The former basin feature is gone (it is now the default).
• The optional nlopt dependency is removed.

Python API

New and improved

• Standard exceptions (#466): invalid input or failure raises ValueError, TypeError, OSError or
  RuntimeError instead of PanicException. Errors raised by your constraint functions are propagated as
  is. Gpx.save() / SparseGpx.save() raise on failure instead of returning False.
• Results: EgorOptim forwards x_opt, y_opt, x_doe, y_doe from its result and unpacks as
  x_opt, y_opt = egor.minimize(...).
• Initial DOE as two arrays: Egor(x_doe=..., y_doe=...), like Egor.suggest().
• Per-constraint tolerance: CstrSpec.leq(bound, tol=1e-3) or {"leq": bound, "tol": 1e-3}.
  When cstr_specs is given, n_cstr is inferred and a mismatching n_cstr raises ValueError.
• Raw constraint layout in results (#475): with cstr_specs, OptimResult.y_opt / OptimResult.y_doe
  (Rust and Python minimize()) hold the constraint values as returned by your function (1 + n_cstr
  columns) instead of the internal <= 0 form (sign-flipped for geq, two columns for eq / between).
  They can be given back to best_result() or Egor(x_doe=..., y_doe=...). DOE files saved in outdir
  keep the internal form.
• best_index() / best_result() (#474) interpret raw constraint values with cstr_specs and their
  tolerances, as the optimizer does. A y_doe not of shape (ns, 1 + n_cstr) raises ValueError.
• Function constraints (#473): minimize(fcstrs=...) also accepts a (g, grad_g) tuple or a
  {"fun": g, "jac": grad_g} dict, with one-argument callables. The g(x, return_grad) form stays.
  Constraints remain g(x) <= 0: a scipy-like "type" key raises ValueError, use fcstr_specs instead.
• Egor(seed=..., verbose=...) gives defaults for minimize() and suggest(); values passed to those
  methods take precedence.
• GP models: Gpx.predict(x, return_std=True) / SparseGpx.predict(x, return_std=True) return
  (mean, std); new nx / ny properties on Gpx and SparseGpx. Input shapes are handled
  consistently across Gpx and SparseGpx (#470).
• sampling(xspecs, n_samples, method=Sampling.LHS, seed=None): method is optional, LHS by default.
• Long enum aliases: ConstraintStrategy.MEAN_CONSTRAINT / UPPER_TRUST_BOUND,
  QEiStrategy.KRIGING_BELIEVER / KRIGING_BELIEVER_LOWER_BOUND / KRIGING_BELIEVER_UPPER_BOUND /
  CONSTANT_LIAR_MINIMUM, FeasibleInfillStrategy.EFI_PROBABILITY / EFI_FEASIBILITY_ENHANCED.
• Type stubs brought in sync with the runtime (#467), and several silent bugs found during the Python API
  review fixed (#468).

Deprecations

Deprecated names still work but emit a DeprecationWarning. Passing both a deprecated keyword and its
replacement raises TypeError.

Deprecated	Use instead
Egor(n_start=...)	Egor(infill_n_start=...)
GpConfig(n_start=...), GpConfig.n_start, gp_config={"n_start": ...}	theta_n_start
GpConfig(max_eval=...), GpConfig.max_eval, gp_config={"max_eval": ...}	theta_max_eval
GpMix / Gpx.builder(n_start=..., max_eval=...)	theta_n_start=..., theta_max_eval=...
SparseGpMix / SparseGpx.builder(n_start=...)	theta_n_start=...
SparseGpMix / SparseGpx.builder(nz=..., z=...)	n_inducing=..., inducing=...
Egor.get_result(x_doe, y_doe)	Egor.best_result(x_doe, y_doe)
Egor.get_result_index(y_doe)	Egor.best_index(y_doe)
Egor(doe=...)	Egor(x_doe=..., y_doe=...)
Egor(cstr_tol=[...])	per-spec tolerance: Egor(cstr_specs=[CstrSpec.leq(0.0, tol=...)]), minimize(fcstr_specs=[...])
CstrSpec.btw(lower, upper), {"btw": (lower, upper)}	CstrSpec.between(lower, upper), {"between": (lower, upper)}
QEiConfig(optmod=...), QEiConfig.optmod, qei_config={"optmod": ...}	optim_every
TregoConfig(n_gl_steps=...), TregoConfig.n_gl_steps, trego={"n_gl_steps": ...}	n_global_local_steps
TregoConfig(d=...), TregoConfig.d, trego={"d": ...}	radius_bounds
sampling(method, xspecs, n_samples, seed) (positional)	sampling(xspecs, n_samples, method=..., seed=...)

Upcoming breaking changes (next release)

• Gpx.thetas(), Gpx.variances() and Gpx.likelihoods() become read-only properties
  (gpx.thetas instead of gpx.thetas()).
• The deprecated names above are removed, and sampling accepts only the
  sampling(xspecs, n_samples, method=..., seed=...) argument order.

Breaking changes

• Previously deprecated APIs are removed (#456):
  • Rust EgorConfig::q_batch(), qei_strategy(), q_optmod() (use the q-EI configuration),
    use_max_proba_of_feasibility(), disable_middlepicker_multistarter(), and the deprecated
    re-exports of SurrogateBuilder / XType (use egobox_moe::SurrogateBuilder / egobox_moe::XType).
  • Runtime flags EGOR_DO_NOT_USE_MIDDLEPICKER_MULTISTARTER and EGOR_USE_MAX_PROBA_OF_FEASIBILITY.
  • Python Egor(outdir=..., warm_start=..., hot_start=...): pass them to minimize().
• The run recorder is removed; state recording is kept for debugging and demos (#455).
• Cargo features: basin removed (now default), nlopt removed, persistent always on;
  new opt-in c-cobyla / c-slsqp.
• Mixed-integer GP mixture refactored (#457): MixintGpMixture handles continuous-only problems without
  overhead and is the only mixture used from Python; load_gp_models is removed.
• With cstr_specs, OptimResult.y_opt / y_doe now use the raw constraint layout (see above).

Bug fixes

• Egor no longer stops at iteration 1 when no initial point is feasible: it falls back to the probability
  of feasibility, then to a random point as a last resort (#465).
• Best point no longer nan when the failsafe strategy is imputation (#464).
• A function constraint may return a one-element array instead of a scalar.

Documentation

• Cookbook recipe and example for constrained engineering problems (lessons from an aero-structural MDO
  study), and extended Egor tuning skill with constraint/failure handling and run diagnosis (#463).
• Docs and typo fixes (#469).

Full Changelog: 0.37.9...0.38.0