chore: Update to Vitest 5 and move hook benchmarks to vitest bench (v6) - #3722

Summary

hooks/benchmark.test.ts fails randomly, and with Vitest 5 as the runner it fails almost every time. It asserts on a single Date.now() measurement:

• the baseline is about 5ms, measured at 1ms resolution
• the baseline is measured once in beforeAll, and the hooked functions are measured later, under different system load
• it runs in parallel with the other 29 test files

This PR updates to Vitest 5 and moves the benchmark to Vitest 5's benchmark API. Benchmark files run in their own project, one at a time and never at the same time as tests. bench.compare() interleaves the iterations of the functions it compares.

 package.json
-  vitest ^3.2.4, @vitest/coverage-v8 ^3.2.4
+  vitest ^5.0.5, @vitest/coverage-v8 ^5.0.5, vite ^7.3.7
+  "bench": "vitest bench --run"
 packages/{feathers,create-feathers}/package.json
-  vitest ^5.0.0, @vitest/coverage-v8 ^5.0.0   (not used by `npm test`, which ran Vitest 3)
 packages/feathers/src/hooks/
-  benchmark.test.ts    Date.now() runtime < 20x baseline
+  benchmark.bench.ts   bench.compare(baseline, hooked): median (p50) overhead < 10x
 vitest.config.ts       benchmark.include, benchmarks excluded from coverage
 packages/feathers/tsconfig.json   benchmarks excluded from lib/ and cjs/
 .github/workflows/nodejs.yml      new "Run benchmarks (non-blocking)" step with continue-on-error

• The assertion compares medians, because single GC pauses make the mean unreliable. Under load the mean ratio reached 52x with an rme of 114%.
• The limit is back to the 10x from the original benchmark in feathersjs/hooks. It had been raised to 20x in v6.
• Vite stays on v7. Vitest 5 allows Vite 8, but Vite 8's Oxc transformer only supports legacy (experimentalDecorators) decorators. Under Vite 8, decorator.test.ts, which uses standard @hooks decorators, fails with SyntaxError: Invalid or unexpected token.

No related issues, and no dependencies on other PRs.

Evidence

Measured as the ratio of hooked to baseline (p50), on Node.js 24.16 and 26.10:

• Before: The old benchmark fails intermittently, for example with Runtime is 226ms, threshold is 160ms. With Vitest 5 it failed 2 out of 2 runs, with Runtime is 208ms, threshold is 180ms.
  After: It passed all 13 local runs, with a maximum ratio of 6.9 against the limit of 10. npm test passes on Node.js 24 and 26 with Test Files 29 passed, Tests 494 passed | 1 skipped. CI passes on 20.x and 24.x, and the benchmark step reports Tests 3 passed on both.

Merge Danger

Door: two-way

Blast Radius: small

• This only affects dev tooling, and the benchmarks are no longer part of npm test. If the benchmark step fails, it shows up in the CI log but does not fail the build.
• Vitest 5 declares node: ^22.12.0 || ^24.0.0 || >=26.0.0. It still runs on Node.js 20 (with EBADENGINE warnings); chore: Require Node.js 24 or later (v6) #3720 removes Node.js 20 from CI.
• Vitest warns that the benchmarks "accessed module export getters too many times". The hooks implementation reads imported bindings through Vitest's module runner, which makes the hooked side slower than it would be in plain ESM, so the check is stricter, not looser. Running the benchmarks against the compiled lib/ would avoid that and could be done in a follow-up.
• Projects that use Vite 8, Rolldown or Oxc can't use the standard @hooks decorators without an extra transform. This is worth mentioning in the v6 docs.

🤖 Generated with Claude Code