Changes in version 0.99.0 Bioconductor pre-submission development version. New features - loo_sensitivity(): sample-wise leave-one-out revalidation of omics differential models with coefficient-shift tracking (delta, standardized deletion shifts, significance flips, and within-term feature-rank changes). - flag_influential(): threshold / MAD / top-k / quantile flagging rules. - profile_flags(): univariate (Fisher / chi-square / Wilcoxon) and multivariable (logistic regression, ORs) profiling of biological and technical sample metadata, including variables in the original model; BH adjustment, expected-count-based test selection, deterministic simulated Fisher tests, missing-data handling, rare-level collapsing, and separation and small-sample safeguards. - influence_by_term(): per-term influence summaries. - Engines: lm, glm (family-aware inference), limma, deseq2 (formula-driven, raw counts, per-fit size-factor and dispersion re-estimation), plus register_engine() for custom engines. - Input adapters: matrix/data.frame always; SummarizedExperiment, ExpressionSet, phyloseq when installed. engine = "auto" resolves deterministically to lm, independent of optional package availability. - future-based parallel execution; results identical across plans under the same seed. Fail-soft per-sample error capture. - Single pseudocount mechanism for log2/CLR transforms. Bug fixes (pre-release review) - CLR transform: corrected column orientation; each sample (column) is now centered across features, preserving the features x samples layout. - Pseudocount: applied exactly once (inside apply_transform()); adapters no longer add their own. - Engine selection: auto now resolves deterministically to lm instead of choosing an engine based on installed optional packages; count-based RNA-seq and normalized expression workflows must select DESeq2 or limma explicitly. - DESeq2 engine: redesigned around the stored model formula and raw integer counts instead of the design matrix; transformed, fractional, negative, or non-finite inputs are rejected rather than rounded; size factors and dispersions are re-estimated for every full and leave-one-out fit. - Feature ranks: added full-data and leave-one-out ranks within each model term, with positive rank changes indicating a weaker numerical rank after sample deletion. - Metadata profiling: original model variables are inferred from the formula, retained in the default profile, and labeled as in-model; simulated Fisher tests use a reproducible local seed without changing the caller RNG state. - glm engine: family-aware dispersion (estimated for gaussian/quasi, fixed at 1 for binomial/poisson) and inference (t-based vs z-based). - Metadata profiling: expected-count-based Fisher selection, reproducible simulated Fisher tests, NA handling, dropped-covariate reporting with reasons, retention and labeling of original model covariates, and separation and small-sample safeguards for the adjusted model. - n_features_shifted counts unique features rather than feature-term combinations.