Changes in version 0.2.1 (2026-08-03)                  

  - spliv_sensitivity_path() now reuses invariant formula parsing,
    complete-case filtering, model matrices, fixed-effect
    residualization, cluster alignment, pattern evaluation, instrument
    scaling, and 2SLS design calculations across its delta grid. This is
    an internal, numerically equivalent optimization: the public API and
    statistical results are not intended to change.
  - spliv() is the sole recommended primary estimator.
  - Fit objects now use class spliv_fit.
  - Accidental legacy aliases present in the 0.1.x interface were
    removed from the public API, including the old
    plausexog()/conley_*() names and implementation helper exports.
  - spliv() now defaults to conventional UCI at delta = 0 when no
    sensitivity method or bound is supplied, and obsolete BPE arguments
    were removed.
  - Confirmatory BPE equivalence intervals are now compared on the
    documented, unit-invariant scale: residual treatment standard
    deviations per one- residual-SD instrument shift. Raw and
    standardized diagnostics are retained separately. This
    corrects 0.1.x eligibility calculations that compared a raw
    coefficient interval with a scale-aware margin.
  - Confirmatory BPE now requires a non-empty transportability rationale
    and uses the aligned sampling or conservative covariance modes in
    standalone validation and final estimation.
  - Exploratory subset diagnostics retain first-stage F-statistics as
    descriptive information but no longer expose an F-threshold
    selection control.
  - Numerical-equivalence, BPE invariance, fixed-effect,
    clustered-inference, edge-case, and plotting tests were expanded. No
    treatment-effect formula or sensitivity estimand was intentionally
    changed apart from the BPE equivalence-scale correction above.
  - A finalized 72-case equivalent-computation benchmark was added under
    dev/. Corresponding sensitivity-interval endpoints agreed to
    numerical precision. The optimized sensitivity-path implementation
    was faster than repeated adjusted-outcome refitting in every
    benchmarked design. Allocated memory was substantially lower in many
    cases but was not uniformly lower, including one large design with
    approximately 25% greater allocation.

                        Changes in version 0.2.0                        

  - GitHub-only public API cleanup release. It was superseded by
    version 0.2.1 before the corresponding CRAN update.

                 Changes in version 0.1.1 (2026-07-29)                  

  - Include generated vignette metadata in source-package builds.
  - Complete initial CRAN-readiness and portability checks.

                        Changes in version 0.1.0                        

Initial public staging release. This version provides uniform and
patterned UCI/LTZ sensitivity analysis, sensitivity paths and tipping
points, fixed-effect residualization, confirmatory BPE
design/validation, and compatibility wrappers for the earlier
plausibly-exogenous API.