Avoid ClassLoader leak from static default async executor (#3178)

[seonwooj0810]

Fixes #3178

Problem

AsyncFeign held its default ExecutorService in a private static singleton (LazyInitializedExecutorService) that is never shut down:

private static class LazyInitializedExecutorService {
  private static final ExecutorService instance =
      Executors.newCachedThreadPool(r -> { Thread t = new Thread(r); t.setDaemon(true); return t; });
}

Inside a servlet container, marking the threads as daemon only lets the JVM exit — it does not help on redeploy. The static cached thread pool retains a strong reference to the initial application's ContextClassLoader, so each redeployment leaks the previous application's ClassLoader and its threads, eventually leading to a Metaspace OutOfMemoryError.

Change

• Replace the static singleton with an instance-scoped default executor created per built client (defaultExecutorService()). Once the built client is discarded, the executor and its (daemon, idle-timing-out) threads become eligible for garbage collection, releasing the ClassLoader.
• Add AsyncBuilder.executorService(ExecutorService) so callers can supply and own a managed, shut-downable executor — the recommended way to control this lifecycle explicitly. It is ignored when a custom client(AsyncClient) is provided.

Default behavior is preserved for callers who don't configure anything: a cached daemon thread pool is still used; it is simply no longer a process-wide static singleton.

Note: feign.kotlin.CoroutineFeign contains the same static-singleton pattern. I kept this PR scoped to AsyncFeign (the reported class); happy to follow up on the Kotlin module if you'd like it addressed here too.

Tests

Added to core/src/test/java/feign/AsyncFeignTest.java:

• usesProvidedExecutorService — builds an AsyncFeign with a spied ExecutorService, performs a request, and verifies the provided executor actually ran the work (submit(...) invoked).
• defaultExecutorServiceStillExecutesRequests — guards that the default (no executor configured) path still works end-to-end after removing the static singleton.

Test evidence

./mvnw -pl core -am test -Dtest=AsyncFeignTest → Tests run: 54, Failures: 0, Errors: 0, Skipped: 0. Code format validated with git-code-format:validate-code-format (BUILD SUCCESS).

Verification done: confirmed no in-flight PR referencing #3178; reproduced the static-singleton pattern still present on master (AsyncFeign.java); maintainer invited a PR with automated tests on the issue; fix touches only .java files; build + targeted tests + format check pass locally on JDK 25.

[velo]

Thanks for the fix and the executor coverage. I cannot approve this while the required CircleCI pr-build check is failing.\n\nGate results:\n1. Test coverage: PASS.\n2. Backwards compatibility: PASS.\n3. Security: PASS.\n\nBlocking issue:\n- Required check ci/circleci: pr-build is currently failing for this PR. Please get the build green before this can be approved or merged.

[seonwooj0810]

I tracked the failing ci/circleci: pr-build down to feign.BaseBuilderTest.checkEnrichTouchesAllAsyncBuilderFields. The PR added the AsyncFeign.AsyncBuilder.executorService field, but the test still expected 12 enrichable fields, while reflection now correctly sees 13.

I pushed ea6c1db to update the expected field count. CI has been retriggered on the new head.

[seonwooj0810]

Follow-up: the first push only exposed the next assertion in the same test. The final fix is 268d8e941055393fbefed2aecc98a83f7fd91ee0: executorService is now excluded from BaseBuilder#getFieldsToEnrich() because it is a caller-owned lifecycle resource and Capability does not enrich ExecutorService. The expected async builder field count remains 12.

[velo]

Thanks for updating the branch. The per-client default executor and caller-managed executor option are covered by regression tests, preserve existing API behavior, and introduce no security concern. Required checks are green.