test!: migrate custom check() suite to tinytest

[davidbudzynski]

Closes #54. Supersedes #64 (testthat 3e approach pivoted to tinytest — rationale in #54 (comment)).

Problem

tests/test_kit.R:7-34 defined check(test,x,y,error,warning) that only did cat("Check ... failed.\n") with no stop()/stopifnot(). Result: R CMD check always green even with regressions (1286 silent checks).

Why tinytest instead of testthat

• kit is lean C code, zero Imports, R >= 3.1.0. tinytest is zero-dependency, R >= 3.0; testthat 3.x is ~15k lines + 22 deps and needs R >= 4.1.
• This suite needs nothing testthat-only (expect_identical / expect_error / tolerant numeric compares; no snapshots, mocking, or reporters).
• Tests ship in inst/tinytest/ and stay runnable post-install via tinytest::test_package("kit") — useful for a compiled package.
• covr is framework-agnostic, so test-coverage.yaml needs no change.

What this PR does

• DESCRIPTION: Suggests: tinytest (drop testthat + Config/testthat/edition)
• New runner tests/tinytest.R (tinytest::test_package("kit"))
• New suite in inst/tinytest/ (29 atomic commits, 1 file each):
  • test-topn.R, test-iif-nif.R, test-parallel.R (psum/pprod/fpmin/fpmax/prange/pall/pany/pmean), test-share.R, plus test-fpos.R, test-setlevels.R, test-vswitch-nswitch.R, test-count.R, test-unique.R, test-psort.R, test-charToFact.R, test-pfirst-plast.R, helper-kit.R
• Mapping:
  • check(id, expr, expected) -> expect_identical(..., info="id") (test_that wrappers unwrapped, legacy IDs as info= labels, function-prefixed for uniqueness)
  • 31 tolerant cases where identical() is FALSE but legacy all.equal()+typeof() is TRUE (complex NA imaginary NA vs 0, C-vs-R floating-point order in pmean/pprod/vswitch) -> expect_kit_equal(..., info="id") helper (mirrors legacy logic via expect_true)
  • check(..., error=) -> expect_error(pattern=, fixed=TRUE, info="id") (mirrors legacy grep fixed=TRUE substring match)
  • 2 warning cases (0001.068, 0020.001) -> expect_warning(..., fixed=TRUE) + value comparison
  • skip_if (shareData) -> exit_file
• Setup and expectations stay interleaved in legacy order (later sections redefine x/y/out_vec, so setup is not hoisted).
• Delete legacy tests/test_kit.R. 17 commented-out psort checks from the legacy file are omitted (were commented).
• test-coverage.yaml: no change needed — covr::codecov() auto-detects tests/tinytest.R.

Verification

• Rscript tests/tinytest.R: All ok, 1271 results, 0 failures (parity with the testthat conversion in test!: migrate custom check() suite to testthat 3e #64).
• Mutation check: 3 deliberately wrong expectations (identical, tolerant-helper, error-pattern) -> all 3 detected with legacy-ID labels. Old suite would only cat() and exit 0.
• R CMD check --no-manual --no-build-vignettes on built tarball: Running 'tinytest.R' ... OK (only pre-existing vignette inst/doc warnings from --no-build-vignettes/missing pandoc).
• Tarball ships inst/tinytest/; tinytest::test_package("kit") on the installed package: 1271/1271 pass.

Notes for reviewers

• helper-kit.R must be loaded with sys.source("helper-kit.R", envir = environment()), not source() — plain source() evaluates in globalenv() where tinytest's recording expect_* aren't visible (noted in the file).
• expect_kit_equal() is intentionally narrow (31 cases). If we prefer strict comparison for those too, we would need to update expected values (e.g. complex NA imaginary) — left as legacy-tolerant for a green baseline.

[davidbudzynski]

All 11 checks green now. Note the latest commit (ci: install knitr from source on macOS) is unrelated to the tinytest migration — it works around an upstream issue: knitr 1.52 (released 2026-09-06) has no usable macOS binary right now (PPM serves a zstd-compressed .tgz that pak can't extract, CRAN mirrors 404). The macOS job failed 3x in dependency setup before any test ran. The step is marked TEMPORARY in the workflow and should be reverted once upstream publishes a readable binary.