fix(summary): handle single fixed-effect row in print.bmmfit summary

[GidonFrischkorn]

Summary

summary() errors when a fitted model has only one population-level regression coefficient. The "Regression Coefficients" block of print.bmmsummary selects rows with sapply() followed by apply(include, 1, any). With a single fixed-effect row, sapply() simplifies to a vector, so apply() receives a dimensionless object and fails:

Regression Coefficients:
Error in apply(include, 1, any) : dim(X) must have a positive length

This replaces that selection with a shape-stable Reduce("|", lapply(...)) (plus drop = FALSE), moved into a small unit-tested helper. Output is identical for models with several coefficient rows.

Cause

select_pars() returns the union of the model parameters and the predicted parameters, e.g. c("dprime", "sdratio"). The printer then marks which rows of x$fixed match any of them:

include <- sapply(paste0(pars_to_print, "_"), function(p) grepl(p, rownames(x$fixed)))
include <- apply(include, 1, any)

When x$fixed has one row, each grepl() returns length 1, so sapply() returns a length-n_pars vector instead of an n_rows x n_pars matrix, and apply(..., 1, any) fails. A single printed parameter collapses the same way.

Fix

.summary_fixed_rows <- function(fixed, pars) {
  patterns <- paste0(pars, "_")
  include <- Reduce(`|`, lapply(patterns, function(p) grepl(p, rownames(fixed))))
  fixed[include, , drop = FALSE]
}

lapply() always returns per-pattern logical vectors of length nrow(fixed); Reduce("|", ...) ORs them into one vector for any row or parameter count; drop = FALSE keeps a single matched row a data frame.

Tests

Two regression tests in tests/testthat/test-summary.R that need no fitted-model fixture (the existing summary test is gated on a real fit, which is why this slipped through): one reproduces the single-row failure, one confirms multi-parameter selection is unchanged.

Relation to the SDT PRs

This is a general bug in core summary.R, so it is a standalone PR off develop rather than part of a model PR. The SDT model stack (sdt_binary ⊂ sdt_mafc ⊂ sdt_ranking ⊂ sdt_rating) is what exposed it. These are the first bmm models that produce a summary with a single population-level coefficient, because they fix nuisance parameters and estimate only dprime. For example, sdt_ranking(dist = "gumbel_min") and sdt_mafc fit as dprime ~ 1 (+ (1 | id)), leaving one estimated coefficient while select_pars() still lists dprime plus the fixed sdratio/mu. Earlier models always had at least two fixed rows, so the flaw never triggered.

The fix is in core code, so every model benefits (any intercept-only single-parameter fit), and the SDT branches will pick it up when rebased onto develop or at merge. They do not need it to build or pass their own tests; they need it only for summary() to print on the gumbel-min ranking and m-AFC examples.