[davidbudzynski]

Summary

This is the 0.0.23 release-review follow-up. It fixes the issues found while checking the package against Seb's review comment: one real performance regression in the new growable duplicate tables, an nswitch threading race, and two long-vector edge cases. Codecov (#6/#67) is intentionally untouched.

CRAN currently has 0.0.21; 0.0.22 was never published, so this branch prepares 0.0.23.

Commits

1. perf(unique): mix hash keys for sequential IDs — src/kit.h, tools/benchmark-unique.R
   The growable duplicate tables took the top bits of a multiplicative hash. Sequential integer IDs mapped to only about 16% of the slots in a power-of-two table, causing long linear-probe scans. Add a cheap xor/multiply mixer before indexing. Equality checks and output order are unchanged.

2. fix(nswitch): make duplicate values deterministic with threads — src/nswitch.c + tests
   The loop over value/output pairs was parallel. With duplicate values, two pairs could write the same output position from different threads. The pair loop is now serial and atomic-output pairs are parallelised over elements instead. Also fixes list-input/logical-output indexing (i instead of j).

3. fix(long vectors): guard logical fduplicated and class topn — src/dup.c, src/topn.c, docs + tests
   fduplicated() used an int loop index for logical vectors and could read out of bounds above INT_MAX elements; it now refuses those inputs with the standard long-vector error. topn() now applies xtfrm() before the long-vector heap path so classed long inputs follow order() semantics.

4. docs(news): prepare 0.0.23 release notes — NEWS.md
   Version header and notes for the fixes above.

Benchmarks

Environment: macOS Sequoia 15.7, Apple silicon, R 4.6.2, single-threaded (kit.nThread = 1, OpenMP disabled for these runs), n = 10,000,000. Times are medians in seconds; 5 runs per case, 7 for the unique cases. “pre-fix” is 0.0.23 before this PR; “0.0.21” is the current CRAN release.

Inputs:

• seq = 100,000 sequential integer IDs repeated to 10m (the case that triggered the regression).
• random = 100,000 random integer IDs repeated to 10m.
• iif / nif / switches as noted below.

Case	0.0.21	0.0.23 pre-fix	this PR	data.table
funique, sequential IDs	0.026	0.604	0.032	uniqueN 0.082
funique, random IDs	0.028	0.051	0.053	uniqueN 0.069
uniqLen, sequential IDs	0.024	0.595	0.032	—
fduplicated, sequential IDs	0.028	0.595	0.032	—
iif	0.016	0.016	0.015	fifelse 0.014
nif, 4 pairs	0.061	0.060	0.060	fcase 0.078
nswitch, 5 pairs	0.019	0.018	0.018	fcase 0.117
vswitch, character	0.130	0.120	0.124	—
vswitch, integer	0.018	0.017	0.017	—
topn(…, 6L)	0.010	0.004	0.004	—
psum, 3 vectors	0.013	0.011	0.011	—

The new hash restores the sequential-ID case from 0.604 s to 0.032 s, and is faster than data.table::uniqueN (0.082 s) for both sequential and random IDs. Sequential funique is now ~1.2x the old full-table version; random is ~1.9x. That remaining gap is the memory tradeoff from #74 (small growable tables instead of a table sized from the input length), and it is offset by the much lower peak memory.

nswitch thread scaling after the race fix (same 10m input, 5 pairs):

Threads	Time
1	0.080 s
2	0.041 s
4	0.024 s

Reproduce the unique case with:

Rscript tools/benchmark-unique.R <library> <label> 10000000 100000

Correctness and checks

• Full tinytest suite with OpenMP enabled (libomp): 1,625 assertions, 0 failures.
• R CMD check --as-cran --no-manual on macOS / R 4.6.2: OK. The only warning is the local checkbashisms tool not being installed; running Debian's checkbashisms against configure and cleanup reports nothing.
• New tests cover:
  • duplicate nswitch values with nThread = 1 vs 2,
  • list-input/logical-output indexing,
  • long logical fduplicated (opt-in long-vector test),
  • classed long topn error path (cheap ALTREP test).
• Existing switch/psum/topn behaviour is unchanged in the benchmark table above.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.84%. Comparing base (99413b6) to head (25a2a55).
⚠️ Report is 134 commits behind head on master.
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #106      +/-   ##
==========================================
- Coverage   98.01%   96.84%   -1.17%     
==========================================
  Files          13       14       +1     
  Lines        5083     5203     +120     
==========================================
+ Hits         4982     5039      +57     
- Misses        101      164      +63     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.