baseline_model() now checks nuisance_list inputs during construction for zero-variance columns, duplicate or near-duplicate columns, non-finite values, nuisance rank deficiency, and columns aliased with baseline terms.
Added nuisance_check = c("warn", "error", "drop", "none") to control whether nuisance problems warn, stop, are dropped with an audit warning, or are skipped.
Added check_nuisance() and clean_nuisance() helpers for inspecting and repairing block-wise nuisance regressors before model construction.
Performance
Sped up event_model() design-matrix construction by replacing the per-term tibble::tibble() / dplyr::bind_rows() calls used to assemble column metadata with a lightweight, validated tibble constructor. This removes the metadata-building hotspot (~15% faster end-to-end on a representative multi-term, multi-run model) with byte-identical design matrices, column names, col_indices/term_spans, and metadata values.
convolve.event_term() now skips columns that are all-zero within a block instead of building and evaluating an empty fmrihrf regressor for each. For block-diagonal-ish designs (trialwise/LSS single-trial models, or factor levels present only in some runs) this is a large speedup (~2.2x faster on a representative 360-column trialwise model) while producing bit-identical output. Designs where every column is populated in every block are unaffected (a fast-exit keeps the original path), and blocks containing NA/NaN fall back to the previous full-column path so filtering semantics are unchanged.
Convolution hot path now shares one fine-grid HRF matrix across all columns and blocks and calls fmrihrf’s C++ evaluator directly, skipping per-column Reg construction / prep_reg_inputs overhead. Combined with a single global output matrix (no per-block zero-alloc + rbind) and deferred tibble materialization in build_event_model_design_matrix(), this is ~2–3.5× faster end-to-end on trialwise/LSS and multi-term workloads while remaining bit-identical to fmrihrf::evaluate(regressor(...)). Per-onset hrf_fun lists and NA-misaligned designs keep the previous path.
Added bench/ cross-library design-matrix benchmarks against nilearn (the FitLins first-level design-matrix hot path). Run with bash bench/run_compare.sh; see bench/RESULTS.md and bench/OPTIMIZATION_NOTES.md.
Bug fixes
convolve_design() now extracts each condition column with dmat[[i]], so it produces correct regressors for base data.frame inputs (its documented example). The previous dmat[, i][[1]] collapsed a data frame column to its first element; the tibble-based internal call path was unaffected.
contrast_weights() and Fcontrasts() for event_model objects now name interaction-term contrasts with the same term tags used by design-matrix col_indices, preventing downstream consumers from dropping crossed-term contrasts because of :/_ key mismatches (#9).
event_model() now warns when continuous parametric modulators are all-zero or have zero variance, catching degenerate design columns before model fitting while preserving the existing design-matrix shape (#8).
Fixed list-based event_model() specifications so hrf(..., subset = ) expressions can use base operators and helper functions from the calling environment, matching formula-interface subset behavior.
Fixed event-model column metadata construction for factor and multi-basis HRF terms that expand to multiple design columns.
Event terms whose subsets select zero events now retain their canonical condition-column names and metadata on the resulting all-zero design matrix, rather than falling back to generic col_1, col_2, … names.
Suppressed exact, known false-positive metadata warnings produced when decorated HRFs are reconstructed by fmrihrf 0.3.0, while continuing to surface unrelated warnings.