-> in packages/pipeline/src/createPipeline.ts:598.

When we enqueue a pending registration, we now materialize a tracked promise (Promise.resolve(maybePending).then(() => undefined)) and immediately attach a .catch(() => {}) before storing it. That tiny catch keeps Node from flagging the rejection as unhandled if the extension’s register() fails before waitForPendingExtensionRegistrations() runs, while the promise itself still rejects so Promise.all(...) inwaitForPendingExtensionRegistrations() surfaces the failure to pipeline.run(). Jest regression suite for the pipeline still passes.

  - Hardened packages/pipeline/src/createPipeline.ts:598 so pending extension registrations stay in the queue until they truly settle. Instead
    of splicing everything up front, we now keep each promise in pendingExtensionRegistrations, remove it in a .finally cleanup (with its own
    guarded catch), and have every pipeline.run() call wait on whatever is still pending. This closes the race where a second run could start
    before earlier async registrations finished.
  - While touching that area I ensured every promise we spin up (the registration result, the tracked pending promise, and the cleanup
    chain) has a defensive .catch(() => {}) so Node never sees unhandled rejections, yet the original error still propagates through
    waitForPendingExtensionRegistrations().
  - The new regression tests in packages/pipeline/src/__tests__/createPipeline.test.ts still pass (pnpm test -- --runTestsByPath packages/
    pipeline/src/__tests__/createPipeline.test.ts).