[billdenney]

Summary

ud_convert() and ud_are_convertible() called into udunits once per element through mapply(). Converting a units vector of n values therefore made n native calls, and each one parsed both units and built a converter. That includes every units<- conversion, which passes a single pair of units. Mixed units had the same per-element cost:

• mixed_units() called set_units() once per element, parsing the unit string for every value.
• set_units() on mixed units grouped its values with keys from as.character(), which pretty-prints every element's units.

Changes

• ud_convert() and ud_are_convertible() convert the whole vector in one call when there is a single pair of units. Otherwise they convert once per distinct pair and write the results back in order.
• Arguments whose lengths mapply() would warn about or refuse, and non-atomic x, still go through mapply(). Results, names, warnings and errors are therefore unchanged.
• mixed_units() builds one units vector per distinct unit and splits it into one units object per value.
• set_units() on mixed units groups with ud_char() keys.
• A NEWS entry.

Behaviour is unchanged

All results are identical() to element-by-element conversion.

• 169 new test expectations compare against the per-element results over an enumerated grid of inputs and 50 seeded random mixes of units.
• They pin order, names (NA-padded when recycled), NA/NaN/Inf, zero-length input, the recycling warning, list x, error cases, and mixed-units construction and conversion.
• The full suite passes: 753 expectations, 0 failures, 0 warnings.
• A deliberately broken version, which writes results back reversed within each group, fails 39 of the equivalence tests, the order tests and three existing tests.

Speed

At 10^6 values:

Operation	Before	After
set_units(), single unit pair	11.3 s	2 ms
x + y	9.6 s	3 ms
ud_convert(), mixed units	8.7 s	134 ms
mixed_units()	4.1 min	6.8 s
set_units() on mixed units	5.6 min	22 s

Every size from 100 values up is faster.

Relationship to other open PRs

Trial merges with #443 and with the long-unit-format fix conflict only in NEWS.md and tests/testthat/test_udunits.R, where each branch appends entries. Keep both sides. R/udunits.R merges cleanly with #443.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.92%. Comparing base (7180a90) to head (1b3f7cd).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #445      +/-   ##
==========================================
+ Coverage   92.56%   92.92%   +0.36%     
==========================================
  Files          20       20              
  Lines        1156     1201      +45     
==========================================
+ Hits         1070     1116      +46     
+ Misses         86       85       -1     

☔ 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.