v0.2.0

v0.2.0 is a correctness-first rewrite. The public surface is one function, bootstrap(X, *, method=...),
configured with a typed, frozen method spec. The 0.1.x class-based API is gone; this is a breaking release.

Added

• One entry point, bootstrap, plus bootstrap_reduce. The latter evaluates a per-replicate statistic
  inside the generation loop and returns only the reduced results, which bounds peak memory regardless of
  replicate count.
• Typed method specs (pydantic, frozen, extra="forbid"): IID; the block family MovingBlock,
  CircularBlock, StationaryBlock, NonOverlappingBlock, TaperedBlock; and the model-based
  ResidualBootstrap (with an AR, ARIMA, or VAR model) and SieveAR.
• Automatic block length: block_length="auto" uses the Politis-White (2004) / Patton-Politis-White
  (2009) selector instead of a fixed sqrt(n).
• Exogenous regressors for the model-based methods: ARX, VARX, and ARIMAX, held fixed during regeneration.
• Uncertainty quantification (tsbootstrap.uq): EnbPIEnsemble, a fit/predict object that retains the
  bootstrap ensemble and produces in-sample and out-of-sample prediction intervals with a choice of
  calibrator (static, sliding-window, ACI, or NexCP); adaptive conformal calibrators aci_halfwidths
  (Gibbs-Candes 2021) and nexcp_quantile (Barber 2023); and forecast_intervals for forward AR simulation.
• DataFrame input through Narwhals: pass pandas, Polars, or PyArrow frames and series.
• sktime / skbase estimator classes under tsbootstrap.adapters.
• An optional compiled VAR kernel (pip install "tsbootstrap[accel]", numba), auto-selected when present.
• A deterministic per-replicate RNG contract: replicate i is bound to its own stream, so results are
  reproducible for a given seed and environment regardless of worker count or chunking.
• diagnose(x) to suggest methods for a series.
• Python 3.13 support (CI matrix and packaging metadata).

Changed

• Recursive residual bootstraps now regenerate replicates from the fitted dynamics and resampled,
  centered innovations (replacing the previous fitted + residuals reconstruction).
• ARIMA replicates are conditioned on the observed initial state with lfilter-consistent residuals.
  Re-injecting a replicate's own residuals now reconstructs the observed series exactly.
• Non-stationary model fits are refused, or skipped via stability_policy, to prevent explosive paths.

Removed (breaking)

• The 0.1.x class API, TSFit, n_jobs/joblib parallelism, the async layer, the feature-flag system,
  and the statistic-preserving method.
• burn_in and initial on the ARIMA spec: these are now rejected at construction, because ARIMA
  conditions on the observed initial state. They remain available on AR, VAR, and SieveAR.

Quality

• Blocking CI gates: mypy (strict-minus) and pyright (strict) at zero errors, ruff lint and format, a
  property-based invariant layer, and a per-file coverage floor.