caverify 0.3.0

• A second exact method, used by default where it fits in memory. Every set of t columns is written as its t-1 smallest columns plus its largest column. For each set of t-1 columns the rows are grouped by their values on those columns, and one pass over each group, one bit per column, decides the coverage of all the later columns at once. On the strength-4 binary arrays used in testing it was about 40 to 50 times faster than 0.2.0 at a few hundred columns, and the gain grows with the number of columns. The method is the one of ca_verify.c (version 2) in ca-tools, described in tools/RESEARCH_ca_verify.md there; this is a portable rewrite of it for R, with mixed levels and the NA rule of 0.2.0.
• The method of 0.2.0 is kept and is used where the new one would need more than getOption("caverify.maxmem", 2^30) bytes or a column has more than 64 symbols. options(caverify.kernel = "direct") or "prefix" forces one of them, for testing and comparison. The result names the method used in a new element kernel.
• Results are unchanged: covered, colsets, gaps and missing_tuples agree with 0.2.0 on every array tested (random uniform, mixed-level and binary arrays with and without NAs, strengths 1 to 6, and covering arrays from ‘CAs’ with a row removed, an entry changed or an entry set to NA).
• The examples are now the first column sets with a gap in lexicographic order, each with its smallest missing value combination, for any number of threads. In 0.2.0 this held for one thread; with several threads the examples chosen depended on scheduling.

caverify 0.2.0

• The default v = NULL now infers per-column symbol counts from each column’s own maximum, identical to v = "auto", which remains as an explicit alias. Versions 0.1.x inferred one uniform value from the global data range, which misjudges mixed-level arrays; changed on a recommendation by Ulrike Groemping, since a default should not assume a uniform CA. Declare a scalar v where the check should also catch a column that fails to reach its intended symbol count.
• NA semantics corrected, a breaking change from 0.1.x. An NA marks a flexible (“don’t care”) entry: a row now contributes nothing to a projection in which it has an NA, so a verified array is covering no matter how its NA entries are later filled. The 0.1.x behaviour let one NA count as every symbol at once, per tuple independently, and could certify arrays that no single choice of values would make covering (a 4-run array passed as a strength-4 covering array on 21 four-level columns). Thanks to Ulrike Groemping for the counterexample. The new behaviour also agrees exactly with CAs::coverage() on arrays with NAs.
• Mixed-level covering arrays (MCAs): v now also accepts an integer vector of length ncol(x) giving each column its own number of symbols, or the string "auto" to infer per-column symbol counts from each column’s own maximum. Tuple counting and indexing in the C kernel are mixed-radix; a uniform array is the degenerate case and takes the same code path.
• A scalar v behaves exactly as in 0.1.x. The only change for existing callers is the v = NULL default described above, which now reads each column’s own symbol count. A uniform array with every symbol present in every column verifies identically under both readings.
• print shows mixed-level results with the levels profile in exponent notation (e.g. levels 4^2 3 2^3).
• Tests: mixed-level brute-force oracle plus randomized mixed-level cross-validation added.

caverify 0.1.3

• Fixed an installation failure on R-devel with clang 22, seen on the CRAN flavour r-devel-linux-x86_64-fedora-clang. The R headers remap match to Rf_match unless R_NO_REMAP is defined, and clang 22’s omp.h uses match as a clause of #pragma omp declare variant, so an omp.h included after the R headers no longer parsed. omp.h is now included first and the C code compiles with R_NO_REMAP and the Rf_ prefixed API. No user-visible change.

caverify 0.1.2

• Automatic thread selection: by default the checker now uses half the machine’s logical cores (single-threaded for small jobs, capped to 2 during CRAN checks). No setup needed; options(caverify.threads = n) or the threads argument still override.

caverify 0.1.1

• Long verifications are now interruptible (Escape / Ctrl-C) without aborting the R session: combinations are processed in batches with an interrupt check between batches, and all working memory is R-managed so interruption cannot leak.
• threads now defaults to getOption("caverify.threads", 1L).
• DESCRIPTION metadata cleanup (URL, BugReports).

caverify 0.1.0

• First version: registered .Call interface, NA as wildcard (“flexible value”), optional OpenMP threading, 0- or 1-based symbol detection, tests including randomized cross-validation against a pure R oracle.