[rent-a-developer]

Closes #53

What does this change?

Adds a performance benchmark suite and a CI gate on allocation regressions. ResXLocalization had no
performance numbers until now. Nothing caught a change that made lookups slower or allocate more, and
consumers had no way to see what the library costs their app.

Refs #53. The spec is the OpenSpec change openspec/changes/add-benchmarks/ (proposal, design, delta spec, tasks).

Benchmarks (benchmarks/, BenchmarkDotNet, .NET 10, JIT only). Each one is compared with the code an
application would write without the library: the .resx designer class and hand-written code over it.

• ResXLocalization.Core.Benchmarks: Localizer.Current.Get with typed, scoped, search-all and enum keys;
  search-all across 1, 5 or 20 resource managers; MessageArguments (FromValues, FromTuples,
  FromObject) against string.Format; and ICU MessageFormat against calling the library directly.
• ResXLocalization.Avalonia.Benchmarks (headless) and ResXLocalization.WPF.Benchmarks (Windows only):
  • The cost per element of {l:Localize}, {l:Localize} with <l:Argument> children and
    {l:LocalizeEnum}, measured on 100 elements, against {x:Static}.
  • Opening 100 or 10,000 elements, against {x:Static}.
  • Switching the culture with 100 or 10,000 live bindings, against a hand-written TranslationSource
    indexer.
  • Every view is compiled XAML shown in a real window.
• Shared/ is compiled into all three suites. Each class first checks that every benchmark produces the
  expected text, then measures.

CI (.github/workflows/ci.yml)

• The three suites run as a matrix in parallel with the build jobs, on PRs and on pushes to main that
  change code. A benchmark-changes job skips them for every other change.
• benchmark-report merges the results (scripts/convert-benchmark-results.ps1) and compares them with
  main using benchmark-action/github-action-benchmark, pinned by SHA:
  • Allocating more than 125% of main's bytes, or allocating anything where main allocated nothing,
    fails the check.
  • A time above 150% of main is an alert only, shown in the job summary and as a PR comment.
  • The label accept-allocation-increase turns the allocation failure into an alert.
• On main, results are pushed to a new benchmark-data branch: the history of the last 200 commits, plus
  the latest reports.
• deploy-docs copies that branch into the Pages site under benchmarks/latest/ and benchmarks/history/.
• The benchmark job gets contents: write and pull-requests: write.

Docs

• New page docs/guides/benchmarks.md, with an entry in docs/toc.yml.
• A "Running the benchmarks" section in CONTRIBUTING.md.
• A row for benchmarks/ in the AGENTS.md layout table.
• A CHANGELOG.md entry under [Unreleased].

No shipped package, public API or runtime behaviour changes. BenchmarkDotNet is referenced only by the
benchmark projects.

Checklist

Always

• pwsh -File scripts/pre-commit-gate.ps1 passes.
• The Release build produces zero warnings.
• Branch name follows <type>/issue-<number>-<slug>: feature/issue-53-add-benchmarks.

If the change adds behaviour or fixes a bug

Does not apply: no runtime behaviour changes. The benchmarks check themselves before they measure.
CHANGELOG.md has an entry anyway, because the results page is user-facing.

If the change touches one UI package

Does not apply: src/ is untouched. The Avalonia and WPF benchmark suites mirror each other.

If the change touches resource lookup, the generated keys, or the packaging that carries them

Does not apply: the only new dependency (BenchmarkDotNet) is referenced by the benchmark projects only, so
no package's dependencies change.

If the change touches the public API

Does not apply.

After merging

• Create the accept-allocation-increase label if it does not exist yet.
• The first push to main creates the benchmark-data branch with the first entry.
• The deployed site serves benchmarks/latest/ and benchmarks/history/.
• The next PR's job summary compares against that entry, with both the Time and Allocations series.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!