                        Changes in version 1.4.0                        

Maintenance

  - Made BLMM reject invalid or ill-conditioned raw anchor Hessians
    instead of ridge-regularising and inverting them. Draw-specific
    Newton updates are now checked against their profiled REML
    objective, and rejected draws fall back selectively to exact lme4
    rather than returning pathological uncertainty.

                 Changes in version 1.3.1 (2026-08-26)                  

Maintenance

  - Registered TMB native routines and disabled dynamic symbol lookup
    for portable compiled-code checks.
  - Made the ALDEx2 compatibility regression test robust to the R 4.7.0
    binomial RNG change by checking overall p-value agreement and
    matching significance decisions instead of a strict elementwise
    tolerance.

                        Changes in version 1.3.0                        

User Facing Changes

  - Harmonized independent scale-uncertainty inputs on standard
    deviations: sample.sm() now accepts s.sd, and coefficient.sm() now
    accepts c.sd.
  - Added explicitly named s.cov and c.cov arguments for covariance
    matrices. The legacy s.var, s.cor, and c.cor arguments retain their
    original semantics with deprecation warnings until ALDEx3 2.0.0.
  - Clarified that gamma is a standard deviation in clr.sm() and
    tss.sm(), and documented that sample-level and coefficient-level
    priors have different covariance structures even when their SD
    inputs are numerically equal.
  - Added a two-group coefficient-scale example explaining how factor
    reference levels determine the sign of a group offset.

New Features

  - Added method = "blmm": an ALDEx3-specific approximate mixed-effects
    engine using a batched profiled REML anchor fit per feature,
    draw-specific local covariance updates, and exact conditional GLS
    fixed-effect solves. Falls back to lme4 for features where the
    approximation cannot be evaluated cleanly.
  - On small datasets (with S\approx 20) blmm is approximately 40x
    faster than lme4, that factor should increase substantially as S
    increases to more realistic levels.
  - Added feature-level parallelism (n.cores) for method = "blmm" and
    method = "lme4".
  - Added dedicated mixed-effects vignette covering model setup, BLMM
    formulation, validation guidance, and runtime comparison with exact
    lme4.

                        Changes in version 1.0.3                        

User Facing Changes

  - Added aldex.plot() for pairwise ALDEx3 model contrasts, with
    volcano, effect, MA, and waterfall plot modes.
  - Added aldex.effect() for ALDEx2-inspired effect diagnostics on a
    single binary ALDEx3 contrast. The function reports the mean
    contrast estimate, mean pooled within-group standard deviation, mean
    Cohen's d, and an ALDEx2-style directional overlap diagnostic.
  - Updated cohensd() documentation and preserved its legacy return
    value: a feature by Monte Carlo sample matrix of Cohen's d draws.
  - Effect diagnostics and MA plots now use reconstructed log abundance,
    logComp + logScale, and clearly report that they require stored
    Monte Carlo arrays.
  - Fixed the Quickstart vignette setup so package data such as
    gut_crohns_data is available during vignette rebuilds.

                        Changes in version 0.1.1                        

User Facing Changes

  - 1000x (approx) speed up in HC3 and HC0 standard error calculations.
    Note this was rate limiting before so this is a huge performance
    boost.
  - Fixed error with how the streamsize variable was calibrated. Leads
    to massive improvement in memory management.
