[tanderson-ld]

What

LDClient.close() tore down the shared executor with shutdownNow(), which discards any already-queued tasks. The terminal OFF data source status is broadcast asynchronously on that same executor during close(), so a registered status listener could miss the final OFF whenever the executor thread happened to be busy at the moment close() ran.

close() now shuts the executor down gracefully (shutdown()), waits up to SHUTDOWN_DRAIN_MILLIS (5s) for already-submitted tasks — including that OFF broadcast — to run, and falls back to shutdownNow() if the drain does not complete promptly so close() cannot hang on a stuck task.

Why

The drop has been latent since the shared single-thread executor was introduced in 5.0.0, but was invisible in practice: before the RETRY changes the meaningful terminal states were real data source failures, not the close-driven OFF. Now that data sources no longer fail permanently and the OpenFeature provider translates a data-source OFF into a provider status transition, a dropped close-OFF leaves a status listener without the terminal state it expects.

The 5s drain cap aligns with the teardown wait LaunchDarkly's Ruby and Python server SDKs already use for their data-system threads. A healthy drain returns in microseconds; the cap only bounds the pathological stuck-task case.

Testing

Adds statusListenerReceivesTheTerminalOffWhenCloseOccursWhileTheExecutorThreadIsBusy, which occupies the single executor thread while the OFF is broadcast on close and asserts the listener still observes it. Verified non-vacuous: it fails on the old shutdownNow() and passes with the drain. Checkstyle clean.

Ticket: SDK-3432

---

Note

Overview
Fixes a shutdown race where LDClient.close() called shutdownNow() on the single-thread shared executor, dropping the queued task that broadcasts the terminal OFF data source status (emitted during dataSystem.close()). Status listeners—and integrations like OpenFeature that map OFF to provider state—could miss the final transition when the executor thread was still busy.

close() now uses shutdown(), waits up to SHUTDOWN_DRAIN_MILLIS (5s) for in-flight/queued tasks to finish, then falls back to shutdownNow() (including on interrupt) so shutdown cannot hang indefinitely. A note documents that any teardown after this step may be delayed by up to that drain window.

Adds regression test statusListenerReceivesTheTerminalOffWhenCloseOccursWhileTheExecutorThreadIsBusy (SDK-3432) with a mock data source that emits OFF on close and a listener that blocks the executor thread during INTERRUPTED.

^{Reviewed by Cursor Bugbot for commit 109e3d6. Bugbot is set up for automated code reviews on this repo. Configure here.}