Changes in version 0.2.0 (2026-09-20)                  

  - 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.

                 Changes in version 0.1.3 (2026-08-23)                  

  - 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.

                 Changes in version 0.1.2 (2026-08-20)                  

  - 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.

                        Changes in version 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).

                        Changes in version 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.